dev-updater/AGENTS.md

# dev-updater

Serves locally-built debug APKs to a phone, and an Android app that
installs them. See `README.md` first for what this is and how to run it;
this file is the working notes on top of that.

The central design point, worth not undoing by accident: **an app is a
project path, not a file path.** Everything downstream — which APK, which
package it replaces, its label, whether it's worth stripping — is derived
from what's actually built under that path, at add time or per request.
Nothing about any particular project is compiled in, and the app list is
mutable at runtime from the phone.

## Layout

- `vendor/wg-app-link/` — a **submodule**, and the shared half of this and
  ai-app: the WireGuard binding, the CA the app pins, QR enrollment,
  owner-only file creation, and the RON house rules. A submodule rather
  than a published crate because it pins an exact commit, so the two
  servers cannot end up on versions of it that disagree. Clone with
  `--recurse-submodules`; `git::pull` runs `submodule update` after the
  merge, because a merge moves the recorded pointer without touching the
  submodule's working tree and the build would otherwise keep compiling
  the old contents.
  What stays here is what differs: this server's routes, registry, build
  walk and service contract, plus `auth.rs`, whose middleware is generic
  over each project's own state -- only the token functions underneath it
  are shared.
- `server/` — Rust + Axum. `main.rs` is the bootstrap and the two
  listeners; `routes.rs` has the whole HTTP table in its module doc
  comment. `auth.rs` is the bearer token every TLS request carries,
  applied once around the whole router so a new route cannot forget it. `registry.rs` owns the live app list and every mutation of it
  (config writes funnel through `AppState::update`, so in-memory and
  on-disk state can't come apart; the one other write is
  `AppState::new`'s startup reconciliation, before anything can read the
  list). `discover.rs` is the scanner,
  `config.rs` the persisted schema and the RON both config files are in,
  `apkinfo.rs` the `aapt2` reads,
  `strip.rs` the slim-APK pipeline, `sdk.rs` the SDK/NDK tool lookups,
  `script.rs` the one runner for a project's short declared commands
  (`resources:`, `enroll:`) -- stdin closed, a deadline, stdout kept.
- `app/` — Kotlin + Compose, a single `:androidApp` module. `UpdaterScreen.kt`
  is the list, `AddAppScreen.kt` the add/settings screen, `AppsApi.kt` the
  management calls, `UpdateManifest.kt` the read side, `ApkInstaller.kt` /
  `InstalledBuilds.kt` the download-and-install path, `DownloadServer.kt`
  the transport and `Link.kt` the two values it hands the shared library. `Theme.kt` is Catppuccin Mocha
  mapped onto Material's roles, plus the four `ActionTone`s buttons come in
  -- red takes something away, the scheme's mauve replaces it with the same
  thing, green brings it up, blue puts a new build on the phone. What a
  button does is said in colour rather than by which component it sits on,
  so the same consequence looks the same everywhere. The mapping that
  matters is the surface ladder: Mocha names its darks in order (Crust,
  Mantle, Base, Surface 0) and Material asks for the same thing under other
  names, so the page is Base, a component's outlined card stays Base beside
  it, and a project's card is Surface 0 -- one visible step, which is all
  the nesting has to say.
  `NerdIcons.kt` names the icon glyphs,
  which are drawn as *text* in a small Nerd Fonts subset rather than
  as vector assets -- an icon beside a line of text wants that line's size,
  colour and baseline, and a `Text` gets all three for free. The font is
  generated: add a codepoint in `NerdIcons.kt` **and** in
  `app/build-icon-font.sh`, then re-run the script, or the glyph silently
  isn't there.

## Checking your work

- Server, from `server/`: `cargo fmt`, `cargo clippy --all-targets`, and
  `../run-tests.sh`. All three every time, not just when a change looks
  big enough to warrant them. The build is warning-clean and
  `cargo fmt --check` passes; keep both true. Formatting is plain rustfmt
  defaults with no `rustfmt.toml` — layout is not something to decide per
  line, so take what it gives rather than hand-formatting against it.
- App: from `app/`, `. ./android-env.sh && ./gradlew :androidApp:compileDebugKotlin`
  (or `:androidApp:assembleDebug`), plus `./gradlew ktfmtFormat
  ktfmtFormatScripts :androidApp:ktfmtFormat :androidApp:ktfmtFormatScripts`
  and `./gradlew :androidApp:lintDebug`. Both are clean; keep them that
  way. ktfmt is `kotlinLangStyle()` with nothing else configured, so
  formatting is never a thing to decide per line.
- Neither replaces running it. `app/run-android.sh` builds, installs and
  launches on an emulator; screenshot with `adb shell screencap -p
  /sdcard/x.png && adb pull /sdcard/x.png <local>`. Package
  `com.example.devupdater`, activity `.MainActivity`.
- **`test-projects/` is what to point it at**, rather than a real project.
  Four fake ones -- a plain APK, one building two variants, one whose build
  fails on demand, and a `Server` + `Apk` pair with a service and a
  `resources:` declaration. They exist because a real project only does what
  it happens to do, where these can be asked to fail, to be slow, or to
  declare nothing. That directory's own README says how they are reached
  (they are deliberately not scanned) and why each one builds into an
  `app/` subdirectory; the short version of the second is that two levels is
  what `find_apks` matches, so a test APK any shallower is offered as a
  build of *this* project.
- The server is easy to exercise directly, which is usually faster than
  going through the UI — but it takes three things, and leaving any of
  them out looks like the server being broken:

  ```sh
  # The CA it actually presents is the one in its config dir, not the
  # stale certs/ left in the repo. The address is wg0's -- it binds that
  # and nothing else, so 127.0.0.1 refuses the connection unless the
  # server was started with --bind. And every route on this port needs
  # the enrolled bearer token.
  curl --cacert "${XDG_CONFIG_HOME:-$HOME/.config}/dev-updater/certs/ca.pem" \
      -H "Authorization: Bearer $TOKEN" \
      "https://$(ip -4 -o addr show wg0 | awk '{print $4}' | cut -d/ -f1):8090/manifest"
  ```

  Only the token's SHA-256 is stored, so there is no reading it back out
  of `config.ron`: either use one printed by `--rotate-token`, or run a
  throwaway server with `--config`/`--certs` pointed somewhere temporary
  and `--bind 127.0.0.1`, which prints a fresh token at startup.

## Things that have bitten

- **Don't add a general recursive search to `discover.rs`.** It was
  measured and rejected; the numbers are in that module's doc comment and
  in the README. Extend `APK_PATTERNS` instead.
- **`aapt2` and the strip pipeline never run on the manifest path.** A
  process spawn or a zip walk there turns a 43 ms refresh into something a
  phone notices, and the manifest is fetched on every open, resume and
  Refresh. `aapt2` is add-time work cached in `config.ron`; stripping
  belongs to the download alone, which is why the manifest reports the
  size of whatever is on disk (`strip::serveable_now`) rather than
  producing the slim copy to measure it.
- **Both config files are RON with two house rules**, and
  `wg_app_link::format` is the only place that knows them -- ai-app's files
  are in the same shape, which is why they are shared rather than described
  twice. `config.rs`'s own `format` module is now only the migration hook
  in front of it, and goes when that does. A file is the *body*
  of the struct -- no outer parentheses, so nothing is indented for a
  wrapper -- because RON has no implicit top-level struct; `parse` adds the
  paren and `render` strips it. And `IMPLICIT_SOME` is set on the
  deserializer rather than by a header each file would have to carry, which
  is why every optional field also needs `skip_serializing_if` so nothing
  is written back that nobody typed. The two halves only round-trip
  together: change one and every file on disk still loads while only
  looking wrong, which is why the config test asserts the written shape.
- **A `Server` component is driven through one script, run as
  `<script> <subcommand>`.** `service.rs` knows the subcommands and the
  four words `status` may print (`running`, `stopped`, `failed`,
  `not-installed`); it knows nothing about systemd or OpenRC, and
  shouldn't. Scripts run with stdin closed, because one that prompts would
  hang rather than fail.
  **On OpenRC, `status` reads the exit code and never the text**
  (`0` started, `3` stopped, `32` crashed, anything else "could not find
  out"). `rc-service status` prints its status line to *stderr*, so the
  obvious check -- discard stderr, grep stdout for `crashed` -- throws away
  the word it is looking for and reports a crashed service as `stopped`,
  which is the exact lie the `failed` state exists to prevent. The codes
  also distinguish "could not find out", which the text cannot: an
  uninitialised user softlevel makes every call fail, and that is not a
  state the service is in. Assign the code through `|| code=$?`, or `set
  -e` kills the script before it can read one -- every answer but "running"
  is a non-zero exit. Measured on OpenRC 0.63.3, not read.
  **There is a guest to measure in**, at `~/vm/gentoo` on this machine --
  `./start.sh`, then `./shell.sh 'cmd'`, ssh on 127.0.0.1:2223 as `tester`
  with `guest_key`. Its own README is the reference. Everything in this
  section, and the restart behaviour below, was run there rather than
  reasoned about; do the same rather than trusting either the
  documentation or a systemd result. Detecting OpenRC is itself a trap:
  `rc-service --user --help` fails when `XDG_RUNTIME_DIR` is unset, so
  probing with it first reports "no service manager here" on a machine
  that has one.
- **Which script is what `service:` says**, and it is a sum:
  `Script("path")` for a project that carries its own, `Managed("the
  binary")` for one that wants dev-updater's built-in. Two variants rather
  than two optional fields, because a component picks one and "both" and
  "neither" would be states the file could express that mean nothing. The
  built-in is `server/src/service-default.sh`, compiled in, written to
  `$XDG_DATA_HOME/dev-updater/service-default` at every startup so an
  updated dev-updater cannot be driving last version's copy, and handed
  its identity as arguments (`--name <key>-<component> --exec "..."`) so
  the subcommand the caller appends still lands last. `service::driver` is
  the single place the two variants become one command -- everything
  downstream takes a command and cannot tell which it was, which is what
  lets a project switch between them without anything else changing.
  It is deliberately a *script* rather than an in-process implementation:
  a built-in that ran through Rust would put systemd and OpenRC knowledge
  back in the one place it has been kept out of, and the contract would
  become a special case of itself. **dev-updater declares `Managed` like
  anything else**, and has no service script of its own: nothing about
  restarting this server lives in a script, it lives in `restart.rs`, so
  there was nothing left for one to say. Its component carries `cwd: ".."`
  because the unit's working directory has to be the checkout root -- that
  is where this server looks for its own app project, and started anywhere
  else its card silently loses the branch line, its commit count and the
  Pull button. `start.sh` knows none of this: it runs
  `dev-updater --service <subcommand>` on the binary it has just built,
  which derives the name and the script from the same declaration the
  running server reads. One place decides how this service is named and
  invoked, which is the only reason a bootstrap script cannot drift away
  from `service::driver`.
  A managed service is named `<key>-<component>`, because a service
  manager's names are one flat namespace across every project on the
  machine. Switching a project from its own script to `Managed` therefore
  changes the unit's name: **uninstall the old one first**, or its unit is
  orphaned with nothing left pointing at it.
- **Service state is checked off the request path**, in `ServiceChecks`
  beside `git::RemoteChecks` and for the same reason: a process spawn per
  server component on `/manifest` is exactly what that path must not do.
  Which means it is subject to the same rule as a remote check: it lands
  *after* the response that started it, so `checksPending` counts both, and
  the phone waits on both (`ManifestEntry.checksOutstanding` for one card).
  And the answering starts at startup rather than on the first request
  (`main.rs`), so the window in which a component has no state is this
  server's own start rather than something a phone has to sit through --
  restarting is how this server is updated, which makes it the moment
  somebody is most likely to be looking at the list.
  Counting only remotes was a real bug and an invisible one: the state is
  unknown only until the first answer of a server's life, remote checks
  ordinarily outlast service ones, and the symptom is a component row with
  no buttons -- so it showed up as "the buttons vanish after a restart if I
  come back to the app too quickly", and only on whichever cards lost the
  race. Nothing looked again until somebody hit Refresh.
- **dev-updater declares itself as an ordinary `Server` + `Apk` pair**, but
  its server component *is this process*, and that changes one thing only:
  *when* it restarts. Running the script's `restart` inside the walk would
  kill it before the APK is built and before anyone is told how it went, so
  the restart is deferred past `finish`; the Restart route defers for the
  same reason, since a script that killed this process before it could
  answer would make a restart that worked read as a failed request. Once
  deferred, both go through the one function (`restart::deferred`), which
  asks the manager when the service is up and only otherwise execs the new
  binary itself. Going through the manager is what makes a refreshed unit
  apply -- an exec inherits the unit this process was started with, so a
  pull that changed the service script would otherwise land only at the
  next restart by hand. The button used to exec unconditionally, on the grounds
  that exec keeps the PID; that made it the one restart a rewritten unit
  never applied to, which is the kind of surprise a special case buys.
  Deciding it in one place is what stops the two drifting again. `BuildState::is_self` is the flag, and it is passed in
  rather than declared -- a file that could claim it is a file that can
  make this server exec an arbitrary binary. A build that changed nothing
  skips the restart altogether (`restart::binary_unchanged`): exec-ing into
  an identical binary drops the phone's connection to deliver the build it
  already had. It is a `stat` of the binary against what it was at startup
  rather than a hash -- cargo installs a new binary by renaming over the
  old one, so the inode alone answers it, and reading tens of megabytes to
  learn the same thing would cost more than the restart it avoids. It says
  "changed" whenever it cannot tell, because a needless restart costs a
  second while a skipped one leaves the old build serving and reads as the
  update having silently failed. The Restart *button* is unconditional --
  somebody asked. Stop and Uninstall do go
  through the script, and do strand the phone; the app confirms them
  rather than hiding them.
- **A component's freshness covers its own directory plus whatever it
  declared with `alsoWatch`** (`Component::watched_paths`, handed to
  `git::subtree_head` and `subtree_dirty` as pathspecs -- git takes
  several, so it stays one call). Scoping to `cwd` is right about where a
  component's files are and wrong about what its build reads: tdep-survey's
  two clients each source one shared `scripts/android-sdk-env.sh`, so a fix
  committed there changed what every build does, moved no component's
  subtree head, rebuilt nothing, and left every card correctly reporting
  "current" -- the narrower-question shape, where nothing is wrong and
  nothing says so. Declared rather than inferred because which files a
  build reads is not knowable from here, and both wrong guesses are
  expensive: too wide rebuilds everything on every commit, too narrow is
  that silence. It is part of the acceptance gate like every other
  declared field -- `same_declaration` destructures exhaustively, which is
  what forced the decision when the field was added.
- **A declaration that cannot be read says so on the card.** RON's
  `deny_unknown_fields` discards the *whole* declaration rather than the
  key it did not recognise, so a machine that takes a project's commit
  before it takes a dev-updater new enough to understand it loses that
  project's components, `strip`, `staleWhen` and resources at once -- and
  the card does not look broken, it looks like a project that declares
  nothing, which is the ordinary case. That was the only difference
  between the two states: a line in this server's log, on a machine the
  person holding the phone cannot read. `config::project_config` now
  answers a `ProjectFile` carrying the parse error beside the
  declaration, `AppEntry::declaration_state` returns both halves from one
  read (the manifest wants both, and the file is on that path), and the
  card draws its own sentence about what it means followed by the
  parser's own words -- selectable, since the fix happens on the other
  machine. Nothing is *blocked* by it: what was already accepted is what
  runs, so an unrelated typo cannot stop a project that was working.
  Worst on the self entry, where `reconcile_self` derives the row from
  that file at every startup: unreadable, and this server's own card
  silently loses its server component and its build commands rather than
  merely stopping noticing changes.
- **The debug keystore is not `~/.android/debug.keystore`** -- that is the
  last of five places, and `sdk::android_prefs_dirs` resolves it the way
  AGP does, because the only thing that makes it the right key is being
  the file Gradle signed with. Measured against AGP 32.3.2 by running real
  builds: `$ANDROID_USER_HOME` is the preferences directory itself with no
  `.android` under it, `$ANDROID_PREFS_ROOT` and `$ANDROID_SDK_HOME` get
  `.android` appended, then `$XDG_CONFIG_HOME/.android` **but only when
  that directory already exists** (the bytecode tests it and falls
  through), then `$HOME/.android`. That fourth step is the whole bug: an
  ordinary desktop Linux machine with `XDG_CONFIG_HOME` set signs its APKs
  with a keystore this server never looked at, so it reported "no debug
  keystore" about a machine that had one and was using it -- and this VM
  sets no `XDG_CONFIG_HOME`, which is why the case could not appear here.
  The failure names a keystore found further down the list, because an
  inert one and an absent one look identical from a phone: mistaking one
  for the other is what turned "this machine has none" into a signature
  mismatch nobody could explain.
- **A stripped copy is only serveable if it is signed by the key the
  build was.** `strip.rs` re-signs with `~/.android/debug.keystore`, and
  nothing about a debug keystore says which builds it made -- it is per
  machine, and one recreated after an APK was built signs as an entirely
  unrelated certificate (measured: a fresh AGP-parameter keystore against
  the existing one, no relation, as a new RSA keypair should be). Where
  they differ Android refuses the package and the phone says "App not
  installed" with no cause, which reads as the download rather than the
  signing and is among the most expensive sentences here to be handed. So
  the two are compared -- `apksigner verify --print-certs` on the source
  and on what was just signed, as sets of digests so a v2 source and a v3
  output still match -- and a mismatch is refused with both certificates
  and the keystore path named. Measured on the files rather than inferred
  from the keystore, so it stays true if the signing step changes. The
  slim copy is deleted on refusal: with no stamp beside it nothing would
  serve it, but `serveable_now` reports the size of whatever slim file is
  on disk, so leaving it would have the card describing a download
  nothing can install. Empty digests mean "not compared" rather than "no
  signer", and say so in the log rather than refusing -- an APK apksigner
  cannot read is one this comparison has no opinion about.
- **A 500 answers with its message.** `ApiError::Internal` used to be a
  bare status with an empty body, on the grounds that an internal cause
  is not safe to hand back -- but every route here is behind the bearer
  token of a device somebody enrolled themselves, so there is no third
  party to withhold it from, and what these actually say is which tool on
  the build machine could not be found or would not run. Withheld, a
  missing NDK reached the phone as "Server returned HTTP 500", with the
  explanation in a log on a machine the person holding it cannot see; the
  app has always shown a failure body when there is one, so the whole fix
  was on this side. The log keeps the same `{err:#}` chain.
- **A remote git failure is shortened and explained in one place**
  (`git::remote_failure`). Git's stderr runs to a paragraph of generic
  advice, so the card gets the first line and the log gets all of it; and
  when the failure was authentication, what this process can see of the
  ssh agent is appended, because "works in my terminal, fails from the
  service" is almost always `SSH_AUTH_SOCK` not being inherited by a
  daemon, and that is invisible from a phone. Asked rather than assumed --
  the note can say the agent is fine, which is what stops it being blamed.
- **A branch with no upstream is not a broken one.** `can_pull` requires an
  upstream as well as a checkout, which is what stops the card reporting
  the same fact twice -- once as its own note and once as the check's
  complaint about it -- and stops it offering a Pull that `git::pull`
  would refuse. It is also not worth *saying*: plenty of checkouts have no
  remote deliberately, and a card that mentions it every time is nagging
  about a choice somebody made.

- **A checkout can be moved from the phone, and it is nearly the same act
  as a pull.** The project settings sheet lists the checkout's branches and
  its last fifty commits, and picking one moves the checkout.
  `BuildState::after_moving` is the half a pull and a checkout share -- one
  copy, because the part that is easy to get wrong is not the git command
  but handing the run over under a single lock, so nothing can observe the
  moment between the move ending and the builds it decided on starting. It
  reports through the project's own state for the same reason: one progress
  path, not a second to keep in step.
  **Where they differ is that a move builds nothing** (`then_build` is
  `None`, which is what `BuildAfterMoving` exists to say). A pull is
  somebody taking new work, so building it is the point; moving the
  checkout is somebody *looking* -- at another branch, at last week's
  commit -- and charging a full build of every component for a look means
  an accidental tap starts a minute of work with no way to cancel it, and
  replaces outputs that were wanted. What it leaves behind is a component
  whose build no longer matches the checkout, which the card already says
  and already carries the button for. Iris asked for this on 2026-09-02.
  Which is why **the build button's word follows the state**: `buildWord`
  in `UpdateManifest.kt` says `Rebuild` only where every component it
  covers is *known* current, and `Build` otherwise -- behind, never built,
  or parked on a commit nobody measured the output against. One rule at
  both scales, so the project row and a component row cannot come to
  disagree about what pressing them means; the project's asks only about
  the components it actually builds, since one with no command has no
  output to be current with anything and counting it would pin that button
  to `Build` for ever.
  The two lists are read by `GET /apps/{key}/refs` when the sheet is
  opened, never on the manifest -- both spawn git, and both are **local
  reads**, so opening the sheet cannot stall on a round trip or fail the
  way a remote check can.
  Two things about listing branches were wrong until a real checkout was
  looked at, and neither shows up in a fixture built to pass:
  `git branch --format` adds a **`(HEAD detached at abc123)` pseudo-entry**
  that is not a branch and cannot be checked out, so the list comes from
  `for-each-ref refs/heads`; and `refs/remotes/origin/HEAD` abbreviates to
  a bare **`origin`**, so the obvious filter for a name ending in `/HEAD`
  matches nothing and a phantom branch called `origin` reaches the phone.
  It is dropped by *being* a symref (`%(symref)` non-empty) instead. The
  test that was meant to cover the second asserted the same wrong thing
  and passed.

- **Fetch is a button, because nothing else here writes a
  remote-tracking ref.** The only `git fetch` this server ran was the one
  inside a pull, and a pull is reachable only when the branch you are
  already on is behind -- so a branch pushed from another machine reached
  the picker only as a side effect of pulling something else, and a project
  already up to date could never be moved onto a new branch at all. The way
  out was a terminal on the build machine, which is where the person
  holding the phone is not. `POST /apps/{key}/fetch` is the fix, sitting
  under the two pickers in the project settings sheet. Deliberately still
  not on a timer and not on opening the sheet: a fetch mutates the checkout
  and pulls down objects, so it stays something somebody pressed, and
  `/refs` stays the local read described above.
  Both pickers sit in a `SettingRow`, which weights the label *and* the
  control rather than letting the control take what it likes: unweighted,
  a pill showing `second-branch-from-elsewhere` squeezes "Branch" into a
  three-character column that wraps one letter per line. The same rule the
  component rows already follow, in the other place a row mixes text with
  a control. The fetch failure and the "reading branches" note go
  *below* both pickers for the same family of reason -- above them, each
  appears and disappears mid-sheet and shoves the pickers down the screen
  as somebody is reaching for one, and the failure ends up nowhere near
  the button that produced it.
  It answers with the **whole ref list** rather than an acknowledgement, so
  the pickers repopulate from after the fetch in the same round trip --
  `routes::read_refs` is the one place both routes build that list.
  `--prune` is asked for here and nowhere else: this is the call whose job
  is to make the picker say what the remote says, where a pull should
  change as little as it can. The phone's read timeout for it is longer
  than the server's own 30s hard stop for a remote command, so a slow
  remote is reported in git's words rather than as the phone giving up on a
  request that is still running. While it runs, both pickers are disabled
  -- it is replacing the lists they are showing -- but uncommitted work is
  *not* a reason to disable it, since a fetch touches no file in the
  working tree. It is the one control in that sheet that still works on a
  dirty checkout.

- **`git checkout origin/topic` detaches HEAD**, which is why
  `git::local_branch_for` exists. Git's DWIM that starts a tracking branch
  fires on the bare name `topic` and on nothing else, and the picker's
  remote-only entries are named `origin/topic` because that is what they
  are -- so picking a branch left the checkout on no branch, with no
  upstream and so no Pull: the state picking a *commit* is meant to
  produce, reached from the control that says it is picking a branch.
  Nothing about it looks like a failure either, since `git checkout`
  prints its detached-HEAD advice and succeeds. Found by running it
  against a real clone once the Fetch button made remote-only branches
  something you could actually reach.

- **The checkout guard is deliberately weaker than the pull guard, and
  that is what stops it being a one-way door.** `git::checkout` refuses
  only on **tracked** modifications (`--untracked-files=no`), where
  `git::pull` refuses on any dirt at all. With the strict check, moving
  back to a commit from before the `.gitignore` that covers this
  project's build output leaves that output sitting there untracked, the
  tree is dirty, and every move afterwards is refused -- you can go back
  and you cannot come forward, from a phone, with the way out being the
  build machine. Nothing is given up by relaxing it, because git makes
  the better check itself: `git checkout` refuses when an untracked file
  would be overwritten and carries across the ones that would not, and
  its refusal arrives as the error the card shows. So this guard covers
  work somebody typed and git's covers the files it would clobber.
  Found by moving a real checkout back and forth rather than by reading
  the code; the first version passed its tests and trapped the checkout
  on the second move.

- **A detached HEAD is a state the card has to say out loud.** Picking a
  commit leaves the checkout on no branch, `git::status` reports the
  branch as the literal `HEAD`, and `can_pull` correctly goes false
  because there is no upstream. The card draws that as **"no branch"**
  rather than `HEAD`, which beside a branch icon reads as a branch
  somebody named HEAD -- worth doing now that the sheet can produce the
  state, where before it was only reachable on the build machine.

- **A checkout parked on a chosen commit is not one to call out of
  date, and the way back is `HEAD`.** Two halves of the same fact.
  The commit list is read from the **branch** (`git::head_branch`, then
  `git log <branch>`), not from HEAD: listed from HEAD, the commits
  *after* the parked one are not in the list, so the picker that moved
  the checkout back could only move it further back -- the same one-way
  door the tracked-only dirty check was written to close, in a place
  that check does not reach. `head_branch` answers the branch a
  checkout is on, or, when it is parked, the single local branch
  containing that commit; `None` when several contain it or none does,
  because there is then no such thing as *the* branch and choosing one
  would be this server deciding which history somebody meant. It rides
  on `/refs` as `head`, and the picker's `HEAD` entry moves there --
  shown as the *current* value while the checkout is following that
  branch, which is what "not pinned to anything" looks like. A checkout
  parked further back than the window gets its own commit appended, so
  the picker never has nothing to show for where it is.
  And `freshness` takes a `parked` flag, reporting a differing commit as
  **unknown** rather than `Behind` while the checkout is detached: every
  way it reads as behind while parked is a build that failed, a
  declaration waiting to be accepted, or a component with no build step,
  and each of those already says so on the same card beside the button
  for it -- so "out of date" there is either redundant or a nag about a
  decision somebody made. Unknown rather than current, because nothing
  measured the output to be what was wanted. `component_is_stale` is
  deliberately *not* changed: Update and Rebuild still rebuild a parked
  checkout, which is the whole point of parking one. The flag is passed
  down from `describe` rather than read per component, since git has
  already been asked for that project's status once and `/manifest` is
  fetched on every open, resume and Refresh.

- **Unrelated histories are the one pull failure the phone may override.**
  A checkout sharing no commit with its upstream has no fast-forward and
  never will, so with nothing offered the card is one that can never be
  pulled again -- and the way out is on the build machine, which is
  exactly where the person holding the phone isn't. So `?force=true` on
  the pull route resets onto the upstream instead of merging, behind a
  dialog that names the branch it is about to overwrite (the only place
  that is seen before it goes). Whether this *is* that failure is decided
  structurally -- `git merge-base` finding no common ancestor -- and never
  by matching what git printed: those messages are translated, and a
  button that appeared only on an English build machine is worse than no
  button. It travels as `PullError::unrelated_histories` rather than
  inside the message for the same reason. The dirty-tree refusal stays in
  front of it, so a forced pull can only ever discard something that was
  committed, and a merely *diverged* history is not offered it -- there is
  something better than throwing that away.

- **The branch line is how a failed remote check is visible at all.** With
  it gone, `newCommits` stays false and the card reads as an app with no
  updates -- the same silent failure the self-entry note below describes.
  It was deleted once by accident in a refactor and nothing failed; only
  the display went quiet.

- **Forcing git onto IPv4 is done in `GIT_SSH_COMMAND`, not on the git
  subcommand.** `git fetch` takes `-4`; `git ls-remote` does not, and
  answers "unknown switch `4'". `ls-remote` is what the new-commits check
  runs, so splicing the flag onto every remote command turned the setting
  into a switch that broke every card's commit count -- which is how it
  was first written, and what `git::ssh_command` and
  `forcing_ipv4_does_not_break_the_check` now hold in place. `fetch` keeps
  the flag as well, since that is what carries the preference to an https
  remote; ssh is covered either way.

- **This server's own project is a row in `config.projects` like every
  other, and everything in that row except `gitIpv4` is re-derived at
  startup** by `registry::reconcile_self`. The row is what gives a
  per-project setting one home -- it used to have to be keyed in a
  separate `config.settings` list, because the one card that is always in
  the list was the one card the setting could not be a field on. What is
  derived stays derived: the path is the working directory (see below),
  and the label and components come from the checkout's own
  `.dev-updater.ron`, so a re-clone or a moved repo fixes itself at the
  next start instead of leaving the row naming a directory nobody pulls.
  Only what somebody chose is carried across. The declaration is accepted
  by construction there, which is why `AppEntry::pending_declaration`
  answers nothing for it: between a pull that rewrites the declaration and
  the restart that follows, the row and the file can differ without that
  meaning anything is waiting to be accepted. A config still carrying the
  old keyed `settings` list loses it silently -- `Config` ignores fields it
  does not know -- so a machine that skips straight past the version that
  migrated one has its `gitIpv4` to set again.

- **A component has two kinds of log, and the kind is an axis of its
  own.** The build log is what this server captured while building it; the
  runtime log is what the component wrote while running, reported by its
  service script. They are asked for one kind at a time (`?kind=`), and the
  modal gives each a tab -- opening on runtime, since what a service is
  doing now is the usual question, and on build for the component a build
  stopped at (`build_failed`). They were one list indexed by generation
  once, and because build logs came first the runtime one sat at an index
  nothing ever asked for: unreachable, and silent about it. ANSI escapes
  are rendered rather than stripped (`AnsiLog.kt`) -- the colour is how a
  process marked its own errors -- and sequences with no meaning on a
  phone are consumed rather than printed, so a cursor movement cannot
  arrive looking like a corrupted log.
- **A project's row has three buttons -- Pull, Update, Remove -- and the
  first two are the same act at two lengths.** Pull takes the commits and
  stops (`?build=false` on the pull route, `BuildState::pull_only`);
  Update pulls, builds what that brought in, and installs every APK it
  produced. Iris asked for exactly this pair on 2026-09-02, replacing
  "Pull & Build" beside a project-wide Rebuild. The point of splitting
  them is that bringing a checkout up to date is cheap and something
  somebody does while looking, where building and installing is the thing
  they press when they want it on the phone -- the same reasoning that
  stopped a checkout building.
  Update is composed in the app out of the two things it already does
  rather than given a route of its own: the pull-and-build reports in the
  project's row exactly as Pull does, and each install reports in its own
  component's row exactly as that component's own Update does. It skips a
  component whose build just failed -- that row already says so, and
  asking again would run the same failing command to say it twice -- and
  it hands the installer **one** APK at a time. An install intent is
  modal and takes the screen, so a project building two APKs would
  otherwise have the second intent replace the first and one component
  would silently never be installed; the rest wait in
  `ComponentState.ReadyToInstall`, which `continuePendingInstalls`
  offers when the screen comes back, through the same machinery and the
  same two triggers as a download waiting for a wrongly-signed copy to be
  removed.
  A component's own install button says **Install** rather than "Build"
  where nothing is built yet: with the right-hand button now saying
  "Build" for a component that is not current, the two read as one word
  meaning two things in the same row. It is named for what you end up
  with, which is also the half it has that the Build beside it does not.
  The project-wide Rebuild is gone with it (each component still has its
  own), and with it `ProjectState.Rebuilding` -- the two remaining
  project-wide states collapsed into one `Working(what, status)` that
  carries its own words, because Pull, Update and moving the checkout
  differ in nothing but the label and the old one said "Pulling and
  building" for all three.

- **"This server has never built it" is something to build, not
  something to sit on** -- and getting that backwards made it the one
  state nothing could clear. `component_is_stale` treated a missing
  `built_from` as no evidence of staleness, so Update built nothing, so
  no build ever wrote the record: a card reading "not built here" with
  Pull, Update and the project's own button all doing nothing about it,
  for ever. Worst for a `Server`, which the output check beside it
  deliberately skips (it has no APK to look for), so nothing else could
  rescue it. Measured on a two-component fixture rather than argued:
  `/prepare` answered `components: []` and the freshness was unchanged
  afterwards; with the fix the same call reports `step: "building"` and
  the card reads `current`.
  So `nothing_built` now asks the record as well as the output, and the
  cost is one rebuild per component on a machine whose records were lost
  -- which was the objection when it was written the other way, and is
  the right price now that the state is *visible*. It is self-clearing:
  the first build writes the record and the question is never asked
  again.
  Which leaves `NotBuiltHere` meaning exactly one thing: a component with
  **no build command**, where nothing here was ever going to measure
  anything and no button would change it. That is the whole difference
  between it and `NeverBuilt` -- one is actionable and flagged, the other
  is quiet, and they were one word covering both.
  `nothing_built` takes the record map rather than reading it: a `lock()`
  temporary inside a call's argument list stays held for the whole call,
  and a test helper doing exactly that deadlocked the moment anything
  under it wanted the same lock.

- **Freshness says *which* answer it is, and only three of them read as
  out of date.** `Freshness` was `current | behind | unknown`, and the
  card drew unknown as nothing at all -- so a component this server had
  never built looked exactly like one whose build was current, and the
  only difference lived in a config file on the other machine. That
  silence cost two rounds of "why isn't it flagging anything" that no
  amount of looking at the phone could answer. It now names the state:
  `behind`, `neverBuilt` and `otherMode` are measured and have something
  to do about them; `uncommitted`, `notBuiltHere`, `noCheckout` and
  `parked` are the ways of not knowing, and each says so in the row in
  the ordinary text colour -- words, not a colour, because "we could not
  check" is a difference in kind from "there is something here".
  `ProjectComponent.isStale` is the one place the division is written
  down, and it decides all three of: whether the row says "out of date",
  whether the card sits above the "Up to date" heading, and whether
  Update is pressable. Iris's rule for it, 2026-09-02: "it should not do
  that until things actually can be updated."

- **Two places rewrite a project's components list, and one type says
  what survives it.** `AppState::approve_declaration` and
  `registry::reconcile_self` both replace `components` wholesale with
  what the checkout declares -- right for everything the project asked
  for, wrong for the three things this machine knows: the package read
  out of an APK, the mode somebody chose here, and the commit and mode a
  build recorded. `CarriedOver` / `carried_over` / `restore_carried_over`
  hold all three, keyed by name *and* `cwd` like every other
  carry-across here, and both callers take the whole thing.
  It is one type because it was three collect-and-restore pairs
  remembered separately, and each caller forgot a different one.
  `reconcile_self` dropped `builtFrom`, so the commit written down by the
  build that produced the new binary was forgotten by the process that
  build started -- and restarting is how this server takes an update, so
  its own card had no freshness ever again: never behind, never current,
  and its Update with nothing to do, since both readings need a commit to
  compare against. Then `approve_declaration` turned out to drop the same
  field, which is worse because it hits *every* project: accepting a
  declaration told every one of that project's cards it had never been
  built here. Silent both times, because "not built here" is a true
  sentence about plenty of projects and so reads as an answer rather than
  as a loss -- it was found by making freshness say which kind of unknown
  it was and then looking at a card that had no business saying that one.
  Adding a field to a component now means adding it here once.
  `reconcile_self` sets the built-in row's package *after* the restore,
  so what is derived wins over an older row carrying something else.

- **One predicate decides both the "Up to date" heading and whether
  Update can be pressed** (`hasWorkWaiting` in `UpdaterScreen.kt`): they
  are the same question, since Update is the button that clears
  everything the heading is about, and a card filed under "Up to date"
  above a pressable Update is the list disagreeing with itself. Three
  ways to have work waiting -- commits on the remote, an APK newer than
  the copy installed here, and **a component whose build no longer
  matches the checkout**. The third was missing, so switching branches
  from the phone left the card sitting under "Up to date" with only the
  small note in one component's row to say otherwise. Iris found it by
  switching ai-app onto another branch and seeing nothing flagged.

- **The list follows a card that an action moved, and a checkout waits
  for the answer before it stops.** Two halves of the same press, both
  Iris's, 2026-09-02. Switching a branch makes the build out of date,
  which moves the card out of "Up to date" and somewhere else entirely
  on a list of ten: `followCard` scrolls to it, but only when it is
  actually off screen, since moving something the reader can already see
  is a jump they did not ask for. It waits two frames first -- the entry
  has just been applied, so the first frame carries the composition and
  the second reports the layout it produced; reading the position before
  both have passed gives where the card *was*. Positions come from an
  `onGloballyPositioned` per card (`CardPlace`) rather than from the
  order, because the cards are different heights and the headings count
  too.
  And `startCheckout` now waits on `awaitCheck` afterwards. The build
  machine drops what it knew about that checkout's remote and asks again,
  but the answer lands *after* the response -- so with a single read the
  card kept a pending check for ever: no commit count, Pull disabled, on
  a branch with commits waiting. That was the whole of "it doesn't show
  any updates or the ability to pull even tho there's obviously new
  stuff", reproduced against a clone of ai-app: immediately after the
  move `newCommits: false, checkPending: true`, three seconds later
  `newCommits: true` -- with nothing asking again in between.

- **A moved checkout re-asks its remote, because the last answer was
  about the branch it left.** `RemoteChecks::recheck` forgets the cached
  answer and starts a fresh check, called from `move_checkout` where the
  thing that invalidated it happens. Without the forgetting the card
  answers for the previous branch with `checkPending` false -- so a
  branch one commit behind its upstream reads as one with nothing to
  pull, and Pull is disabled about it with nothing on screen saying the
  answer predates the move. Measured, not argued: git said `[behind 1]`
  while the card said `newCommits: false`. `Checks::forget` is the
  primitive, and it deliberately leaves `in_flight` alone for the reason
  `update` does -- a worker already running still owns the right to write
  its answer.

- **There are three sizes of refresh, and using the wrong one is what
  makes the list feel like it has a mind of its own.** `refresh()` drops
  the list to a spinner and asks every remote -- it belongs to arriving,
  resuming, and the Refresh button, which are the moments somebody asked
  to be shown the current state. Pull-to-refresh is the same ask without
  the spinner, since the gesture brings its own indicator and taking the
  list away underneath it would say the same thing twice. `applyOne` reads `GET /apps/{key}`
  and puts that one entry back, so an action on one card cannot move or
  change any other; every card action uses it, including adding one --
  the Add screen hands back the key and the list fetches just that app.
  `awaitCheck` is the middle one: ask a single checkout's remote and wait
  on *that* answer, which the card's own refresh control and a freshly
  added app both use. Waiting on the whole list to settle would make one
  card's refresh sit behind another card's.
  Arriving and resuming are **one** event, not two: `LifecycleResumeEffect`
  runs when the screen first reaches RESUMED, so there is no
  `LaunchedEffect(Unit)` beside it and one thing decides when the list is
  read. It refreshes from *any* state including a failed one -- guarded on
  "loaded", as it used to be, a card that had gone red stayed red until
  somebody found the Refresh button, which is the opposite of what
  returning to an app should do. `loadingList` is what stops the first
  composition and the first resume stacking two reads, and it is claimed
  *before* the coroutine launches, because both run in the same frame and
  a flag set inside the coroutine is set too late to be a guard.

- **A failure that lands while nobody is looking is not shown.** Work
  started before the app went away keeps running -- deliberately, since a
  download that finishes in the background is a download that worked -- and
  when the device sleeps or the link drops it fails. Reported, that meant
  coming back an hour later to a five-second read timeout that said nothing
  about the server and that there was nothing left to do about. So every
  catch in `UpdaterScreen.kt` goes through `failure(e)`, which answers null
  while the screen is not resumed, and **null means clear, never leave** at
  every site -- a dropped failure that left the card alone would leave a
  spinner up for an operation that has already stopped. `setProject` and
  `setComponent` both take a nullable state so the answer can be handed
  straight over. Iris asked for this on 2026-09-01: "I got a socket timeout
  by leaving the app for too long. Make sure not to show that if the app
  just gets unloaded."
  The other half of the same complaint is the read on the way back *in*:
  the link may have been asleep as long as the app was, and the first
  request across one still coming back times out at the five seconds every
  request gets. So the resume's read alone retries once (`afterAGap`),
  which costs nothing when the server really is down -- a refused
  connection comes back at once rather than waiting out a timeout. That is
  a retry, not a guess: nothing is displayed that was not measured.
  Dropping a failure is only safe *because* the resume re-reads from any
  state; the two changes hold each other up, and undoing either alone
  leaves the list stuck on a spinner or stuck on a stale error.

- **The Add screen is drawn over the list, not in place of it.** Swapped
  out, the list is composed again from nothing on the way back -- and
  rebuilding it means fetching it, which is the whole-list refresh that
  adding an app has no business causing. `describe` in `routes.rs` is the
  one place a card is built, so `/manifest` and `GET /apps/{key}` cannot
  come to describe the same app differently.

- **A progress bar is drawn from counts the build reports, never from an
  estimate**, and the knowledge of how to get those counts belongs to this
  server rather than to each project. Two ways in, both ending as
  `@@progress done/total` on the build's output:
  cargo reports its own once `CARGO_TERM_PROGRESS_WHEN=always` is set,
  which `spawn` sets for every build command, so a cargo build needs
  nothing and must not be wrapped. Gradle cannot -- an init script is
  refused by the configuration cache, and its rich console reports a
  percentage only as a full-screen redraw that would make the output
  unreadable -- so `server/src/build-progress.sh` does the `--dry-run`
  count and the `> Task` tally, shipped by this server (`crate::shipped`)
  and offered to build scripts as `$DEV_UPDATER_PROGRESS`. A script tests
  for the variable rather than depending on it, so building by hand still
  works. This is what stops each project carrying its own copy of the
  same twenty-five lines.
  Reading the output means splitting on carriage returns as well as
  newlines: a tool redrawing a counter in place puts several updates and
  then real output inside one `\n`-delimited line.
- **Discovery must stay side-effect free.** It runs on every manifest
  request and every suggestion scan.
- **An app in the list is a *project*, and what it produces is its
  components.** Ordinarily one `Apk`; a project that also runs a server on
  the build machine declares a `Server` beside it. They are a sum, not one
  struct with both halves' fields, and the common fields are repeated per
  variant so the file reads `Apk(name: ...)` with nothing nested inside a
  `kind`. `Component::same_declaration` destructures exhaustively, so
  adding a field anywhere fails to compile until someone has said whether a
  project declares it or this server measures it.
- **A build reinstalls every server component's unit, not just changed
  ones.** The unit is generated from a script in the repository, so a pull
  can change how a service is *defined* rather than what it runs -- and
  that has to land even when the build produced a byte-identical binary,
  which is the ordinary case for a pull that only touched the script.
  Installing is safe on a running service (it writes the unit and reloads;
  it stops nothing), which is why `start.sh` has always done it every run.
  It is deliberately not gated on `restart::binary_unchanged`: that answers
  whether a *restart* is worth the interruption, and unit staleness is a
  different question. Wiring the two together left the unit stale exactly
  when the build had nothing to do.
- **A build runs every component at once, and each reports for itself.**
  They are independent -- a Rust build and a Gradle build share nothing but
  the machine -- and measured here, together takes about three quarters of
  the time one after the other does. The saving only appears when more than
  one has work, which is what a pull produces.
  The status is per component (`ComponentStatus`), so a card draws each
  component's bar and last line inside that component's own row;
  the project's own area at the bottom keeps only what belongs to the whole
  project, which is fetching and pulling. A bar under the card could only
  ever say that *something* was happening, and with everything building at
  once that is exactly what the reader is trying to find out.
  A failure no longer stops the others -- they are already running -- and
  it is reported against the component whose command it was, never as the
  project's. The phone draws every component in
  a card of its own, including a project with only one: the flat layout it
  used to get meant two shapes to keep in step, and put that APK's size up
  beside the card's corner controls where it read as belonging to them.
- **There is a build slot per component, not per project, and the two
  halves have to agree on that.** A component being built neither blocks
  another's build nor disables its controls: `Inner` has no `building`
  flag, only a `ComponentRun` per component whose open `step` *is* the
  answer, and `claim` writes that entry synchronously under the same lock
  the route answers from -- so nothing can read a component the request
  just claimed as idle, which the phone would take for "the build is
  over". `BuildStatus::building` stays, but it means "anything at all is
  happening here" and is only for the controls that act on the whole
  checkout; anything about one component reads that component's `step`.
  The app mirrors the split exactly: `ProjectState` for the pull and the
  project-wide Rebuild, `ComponentState` keyed by component name for
  everything a single component is asked to do. One map keyed by project
  alone is what the bug was -- pressing Update on one client of a
  two-client project disabled the other's button and drew this one's
  download bar under it -- and two hierarchies rather than one keyed by a
  pair is what stops it coming back, since a download has no
  project-wide meaning to be stored with.
  The exception, and it is worth keeping visible so it does not read as
  more of the same: **a pull really is exclusive with everything.** There
  is one checkout, and it rewrites the files every component builds from,
  so `Inner::pulling` blocks any component from being claimed and waits
  for any still building. Ending the pull and claiming what it decided to
  build happen under one lock for the same reason `claim` is
  synchronous -- a phone polling in the gap would see a project that is
  neither pulling nor building and call the run finished.
- **Text a command produced is selectable; text this app wrote is not.**
  `Theme.kt`'s `OutputText` is the whole of it, and every failure message
  goes through it -- a component's build or download, a service action, a
  checkout's remote check, and the log dialog's own failure line -- plus a
  `SelectionContainer` around the log body, which cannot use `OutputText`
  because it is an `AnnotatedString` the ANSI renderer coloured inside its
  own scrolling panel. The reason is that this is the one text on screen a
  person has to take somewhere else, and the machine that produced it is
  not the machine in their hand. A status word or a button label stays
  unselectable on purpose: selection handles on those are noise, and a
  card that starts a selection on long-press fights the gestures it
  already has. Iris asked for exactly that line on 2026-09-01: "not the
  'failed' but the command output for build errors and stuff".
  `OutputText` also renders the ANSI escapes rather than printing them,
  through the same `ansiAnnotated` the log dialog uses -- a compiler marks
  its own errors in colour and the tail of a failed build is that output
  verbatim, so raw it arrived as `[1;31merror` with punctuation welded
  onto the one line somebody was trying to read. Selection copies
  `AnnotatedString.text`, which is the message with every escape already
  gone, so what lands on the clipboard is what was on screen rather than
  what was on the wire.
- **A finished component shows nothing, and its button goes back to
  normal.** Iris's call, 2026-09-01: "you shouldn't see the time it took
  once it finishes, it should just go back to its normal enabled button
  state." So `ComponentBuildProgress` draws only while the step is open,
  and the elapsed times are gone from both halves of the wire -- a bar, a
  count and a last line all describe something happening *now*, and every
  one of them sits there looking live beside a sibling that genuinely is.
  What says the build landed is the control becoming pressable again and
  the card's own freshness. The failure is the exception, because it is an
  outcome rather than residue; it is drawn by the component card next to
  the Retry that acts on it, in the one place that reports that
  component's failures whether they came from a build, a download or a
  service action.
- **A card being worked on has no freshness, rather than the one from
  before the press.** Nothing re-reads the manifest during a run -- the
  entry was fetched before the button was pressed and read again only
  once the run is over -- so "out of date" drawn beside the bar that is
  making it current is last minute's answer wearing this minute's
  clothes. `UpdaterScreen` therefore maps a busy component's `freshness`
  to `unknown` in the one place the component list is built for the
  cards, which is what makes both readers of it -- the row's own note and
  `MismatchedPairNote` -- go quiet without either having to know why. The
  condition is deliberately the same pair that disables the Update button
  (`projectState.busy || componentState.busy`): what cannot be acted on
  is exactly what cannot be measured just now. It comes back the instant
  the run ends, still saying "out of date" if the build failed, because
  by then the entry has been read again -- withheld is not the same as
  cleared.
- **A project's own `.dev-updater.ron` is a request, never an
  instruction.** It only runs once accepted from the phone, which copies
  it into `config.ron`; `AppEntry::pending_declaration` is the whole gate.
  Evaluate it *fresh* rather than caching it on the entry — the file
  changes on a pull, and a pull writes no config, so nothing rebuilds the
  entry list. Caching it meant a pull could swap the command out from
  under a previous acceptance and nothing noticed until an unrelated
  config write. Enforce it wherever a command would actually run, not
  only where the list is built.
  Two things follow that both went wrong the moment `resources:` joined
  the gate. **Whatever `Declaration::matches_accepted` compares,
  `AppState::approve_declaration` has to write** -- it compared the
  resources declaration and wrote only the components, so accepting stored
  half of what was being read back and the gate could never clear. The
  press succeeded, so there was no error either; the card simply kept
  asking. Adding a field to one without the other is the shape to watch
  for, and the tests are what let it through: they built an "accepted"
  config by hand instead of calling `approve_declaration`, so they
  asserted a copy of the rule rather than the rule.
  And **on the pull path the gate is asked after the fast-forward, never
  before it** (`build_state::pull_and_build` takes a closure for exactly
  this reason). The declaration lives in the checkout, so the pull is the
  one thing that changes the answer -- a bool computed at the call site is
  the answer for the commit being replaced, which made a pull that changed
  the declaration build anyway. Pulling stays allowed while a request is
  unaccepted, since taking commits runs git rather than the project's
  command, and it is how the new request arrives to be read.
- **A component declares its build modes in one list, and that list is the
  only place a mode is declared.** `modes: ["release", "debug"]` on the
  component; the first is the default, which is why a project puts the one
  it wants built by default first. Every command the component runs is
  then handed the mode **as its last argument** -- so a project whose
  script takes `release` or `debug` names that script once and nothing is
  written twice. The escape hatch is writing a field *per mode*
  (`build: {"release": ..., "debug": ...}`, `service: Managed({...})`),
  and it exists because some commands cannot take the word: cargo takes
  `--release` or nothing, and its profile for the unoptimised build is
  called `dev` while the directory it writes is called `debug`, so no
  single word serves as both the flag and the path. Gradle is the same
  shape (`assembleRelease` needs capitalising).
  **A command written per mode is not handed the mode as well** -- it
  already is the answer, and appending the word would pass a stray
  argument to a binary that never asked for one, which for a service is a
  process that will not start. `ByMode::mode_argument` is the one place
  that rule lives.
  Having *one* list is what makes "this part has a mode the other part
  never heard of" unsayable rather than something to detect: every
  per-mode map is checked against it (`config::mode_problem`), so a
  missing entry is named as missing instead of resolving to some other
  mode's command. A map on a component that declares no modes is the same
  mistake from the other side and is refused with a message saying what
  to write. A declaration that fails the check is discarded exactly as an
  unknown field is, and the card says so.
  **`ByMode` holds a `Command` and nothing else, and that is load-bearing
  rather than incidental.** Telling the single form from the map needs
  `deserialize_any`, and RON reports a struct `(path: "a")` as a *map* --
  so a struct-valued field behind it would read its own field names as
  mode names. Worse, measured: RON **discards the variant name** under
  `deserialize_any`, so `Cargo("server")` and the array command
  `["server"]` arrive as the same one-element sequence. That is why the
  by-mode map lives *inside* `Service` rather than around it, and why a
  `Cargo(...)`-style shorthand cannot be added to `build:` beside the
  bare-string form without a format migration.

- **Which mode to build in, and whether to strip, are the build
  machine's choices; which finished build to install is the phone's.**
  There is one checkout and one set of outputs, so a per-device mode
  would have two enrolled phones rebuilding over each other with nothing
  on either screen to say why -- the settings sheet says as much out
  loud. They are stored on the component in `config.ron` (`mode`,
  `stripHere`), set through `PUT /apps/{key}/components/{name}/settings`,
  and excluded from `Component::same_declaration`: choosing one of the
  declared modes is not the project asking for something new, and
  including it would make every settings change re-open the acceptance
  gate. Which means both halves of the carry-across have to be written --
  `registry::chosen_settings`/`restore_chosen_settings`, called by
  `approve_declaration` **and** `reconcile_self`. Without the second, a
  mode chosen for this server's own component would last until the next
  restart, and restarting is how this server is updated.
  `strip` stays declarable by the project and `stripHere` overrides it,
  rather than being copied in at acceptance: copied, a later change to
  the declaration would be ignored on every machine, silently.
  **`BuildState::matches` compares the effective mode as well as the
  declaration**, because that state holds a *snapshot* of the components
  it builds from -- reused across a mode change it would go on running
  the old mode's command from a card reporting the new one.

- **A mode switch moves no commit, so `builtMode` is recorded beside
  `builtFrom`.** Without it a component built in `debug` and switched to
  `release` reads as current, offers nothing to press, and serves the
  debug build for ever -- and for an `Apk` nothing else would notice,
  because the "never built at all" check finds *any* variant under the
  component's directory, so the debug APK sitting there is enough to
  satisfy it. `Component::built_in_another_mode` answers it and both
  `freshness` and `component_is_stale` ask, the first reporting `Behind`
  *before* consulting the checkout at all: a clean tree at the very
  commit the debug build was made from still does not make that build a
  release one. It stays quiet until this server has built the component
  once, so a project built by hand is not told it is out of date.

- **Each component card has a settings sheet, and there is one picker in
  it, not two.** The gear sits at the row's right-hand end beside the log
  button, drawn unconditionally so its presence is never the signal. The
  sheet holds one **Build** choice, whether to strip, and Enrol; a
  section with nothing to offer says so rather than vanishing, since
  "this project declares one way of building" and "we could not tell"
  must not look alike. Everything applies on Save except Enrol, which is
  an action and happens on the press.
  The build mode and the installed variant were two dropdowns, and Iris
  asked for one (2026-09-02). They are genuinely different questions --
  what the build machine *builds*, and which finished build this phone
  *takes* -- but they answer to the same words, so two pickers both
  offering `debug` and `release` read as one choice asked twice; moving
  the variant off the card into the sheet had not fixed that, it had only
  moved it. So **where a component declares modes, the mode is the whole
  answer**: `AppEntry::resolve_apk` serves the build named after the
  component's effective mode when the phone asks for no particular one,
  which is what lets the phone stop having a second opinion. Saving in
  that state clears any stored variant outright rather than leaving it,
  since a pin set before the modes existed would otherwise go on quietly
  overriding a control that is no longer shown. A component with no
  declared modes still gets the variant list, because otherwise there
  would be nothing to choose at all -- and the note says which of the two
  it is offering, so "the machine decides" and "this phone decides" are
  never confused.
  Serving the mode's build rather than the newest is the right default
  for its own sake: a debug build made by hand after the mode was set to
  release is *newer*, and serving it would put a debug app on the phone
  from a card saying release.
  Every dropdown in the app hangs from one `PickerButton` -- an outlined
  pill with a chevron, Iris's ask in the same message -- so "there is a
  choice here" looks the same for a mode, a branch and a commit. The
  chevron is `md-chevron_down`, which like every other glyph had to be
  added to **both** `NerdIcons.kt` and `build-icon-font.sh` and the font
  regenerated.
  **The row's text is inside one weighted child so it can never push the
  controls off the edge**, and every reading in it truncates with an
  ellipsis. A control that leaves because the text grew is one the reader
  cannot get back to.

- **`enroll:` is a command whose one line of stdout is a URL for the
  phone to open after installing.** Deliberately "a URL to open" rather
  than anything named after enrolment -- a route that knew what enrolling
  was would be a special case of itself. Run **per press**
  (`POST /apps/{key}/components/{name}/enroll-link`), never cached: the
  link a project mints is ordinarily one-shot and carries a credential,
  so a stored one would be both stale and a secret sitting in a file. In
  the acceptance gate like `build`, because it runs on the build machine.
  `crate::script::capture` is the shared runner it and `resources` both
  use -- stdin closed, a deadline, stdout kept, and a failure naming the
  command and the first line of its stderr.

- **Which build variant to serve is the phone's choice, not the server's.**
  It arrives as `?variant=` on the download, beside the `?component=` that
  says whose build it is, and is validated against *that component's*
  discovered builds -- an unvalidated path would let a phone name any file
  on disk to be served. Storing it server-side meant one enrolled device
  silently changing what another was offered. The stored choice is keyed
  by project *and* component on the device, so pinning one client to a
  release build says nothing about the other. Naming none is answered by
  the component's own mode before it is answered by the mtime; see the
  settings sheet below.

- **A build that cannot be installed over what is on the phone says so
  before the installer does.** Android refuses a package signed by a
  different key than the installed copy and reports it as "App not
  installed" with no cause, which reads as the download having failed.
  So `SigningKeys.kt` compares the downloaded APK's certificates against
  the installed package's, after the download and before the install
  intent, and the card puts up a dialog naming both digests with the one
  thing that gets past it: remove the old app. Iris asked for this on
  2026-09-02.
  It is **not** a guarantee and must not become one: where either side
  cannot be read the answer is "don't know", the install goes ahead, and
  Android decides exactly as before -- blocking on a guess would be worse
  than the sentence it replaces. Certificate *sets* are compared first,
  because a multiply-signed package is only replaceable by one signed by
  all the same keys and `hasSigningCertificate` reports false for it
  however the certificate is presented; that call is then asked as well,
  because it is the only thing that knows a rotated key's lineage.
  The download is carried on the state (`ComponentState.WrongKey`) so
  that removing the old app is followed by the install it was for rather
  than by a second download. The removal happens in the system's own
  dialog with this app off screen, so `continuePendingInstalls` is called
  from **both** things that can learn it finished -- the package
  broadcast and the resume -- since which arrives first depends on how
  long somebody spends in that dialog.
  **`ACTION_DELETE` needs `REQUEST_DELETE_PACKAGES` now**, and without it
  the failure is invisible from this side: the uninstaller starts, logs
  that this uid lacks the permission, and finishes without drawing, so
  the button reads as dead. It needed no permission when this app's first
  uninstall offer (the renamed-package one) was written, so that button
  had presumably never worked and nothing said so. Found by pressing it.

- **A project can produce more than one APK, and each component's builds
  are found under its own `cwd`.** `APK_PATTERNS` is anchored at
  `project.join(cwd)` rather than at the project root, which is what
  `cwd` already meant everywhere else -- the directory the build command
  runs in, the subtree `subtree_head` scopes staleness to, a server's
  `WorkingDirectory`. A component that declares none sits at the root,
  which is what every single-APK project has always meant, so nothing
  about that case changed. tdep-survey is the project that needed it: one
  checkout, a backend and *two* independent Android clients, where the
  root-anchored patterns reached the first and stopped.
  The alternative -- naming the file on the component, `apk: "path"` --
  was rejected because it makes a component a file path, and **an app
  being a project path rather than a file path** is the invariant at the
  top of this document.
  Everything downstream is per component in consequence: `package`,
  `strip`, the variant list, the size, the mtime and the rename note.
  Those were all `apk_component()`'s *first* match before, which was
  correct only while a project had one. Two clients install over
  different packages and differ in whether their symbols are worth
  carrying to a phone, so first-wins would have checked the installed
  state of one app and reported it as the other's -- the expensive kind
  of wrong, because it looks exactly like an answer.
  **A download that names no component is refused, not guessed**
  (`ApiError::AmbiguousApk`), for that reason. Naming none still answers
  for a project with one, which is nearly all of them and is what lets
  the frozen `/self/apk` keep working -- it cannot carry a component, and
  the project it describes has a single APK.
  The measurement carried across an acceptance is keyed by name *and*
  `cwd` (`registry::component_id`): once the directory decides which
  builds a component has, a reused name is not the same APK, and handing
  it the old one's package would be wrong until that component happened
  to be downloaded.
- **`/prepare` and `/build` take `?component=` to build one component
  instead of every one with a command.** Without it, pressing Update on
  one client of a multi-client project ran every component's build to
  get the one that was actually asked for -- fine when a project had one
  APK, expensive the moment it had two and one of them was slow.
  `named_component` in `routes.rs` is the one place a name from the phone
  is checked against the project's own components, so `/prepare` and
  `/build` cannot disagree about what an unknown name means, and
  `BuildState::{trigger_if_needed,build_now,run_build,is_stale}` all take
  the same `Option<&str>` -- `None` still means the whole project, which
  is what `Pull & Build`, the project-row `Rebuild`, and every project
  with a single component keep doing. **Each component's row carries its
  own Rebuild**, which is `/build` scoped to it. That used to be
  deliberately absent, on the grounds that Update already builds a
  component that needs it -- but Update runs `/prepare`, which builds only
  when the staleness rules say the output is behind, and those rules are
  blind to everything a commit does not describe: a command reading files
  nobody declared, an output changed underneath this server, a signing key
  replaced since the APK was made. In all of them nothing reads as stale,
  so Update does nothing, and the only force was a project-wide Rebuild
  that rebuilds every sibling too. Iris asked for the per-component
  button on 2026-09-01. The two are the same action at two scales, so
  they share a word and a colour, and each sits at the right-hand end of
  its own row.
  A component with no command of its own draws no Rebuild -- `hasBuild`
  on the manifest component says so per component, which the project's
  own `needsBuild` (the OR of every component's) cannot. Absent rather
  than disabled, matching the project row, because having no build step
  is not a state a component is in.
  A server's Rebuild sits in its service row but outside the condition
  that draws the service buttons: those wait for the script's answer,
  while a server whose script cannot be reached is still one this machine
  can compile.
  **Where `component.dir()` belongs is now one definition**
  (`Component::dir` in `config.rs`), because a second one very nearly
  shipped a real bug: `component_is_stale`'s "never built at all" check
  still asked `find_apks` of the *project root* after per-component
  discovery had already moved everywhere else to `component.dir()`. A
  project with two `Apk` components has one's output sitting under the
  root-anchored patterns too -- `*/build/outputs/apk/*/*.apk` matches any
  one-level subdirectory, regardless of which component put it there --
  so the moment *either* component had ever been built, the whole
  project read as "something is built here," and the *other* component,
  never built, silently stopped being offered its own first build:
  `prepare` saw a component with a command and no output and declared it
  current. Caught by testing the actual behaviour of a two-APK project
  rather than trusting that scoping the build implied scoping the
  staleness check that decides whether to run it -- they are two
  different reads of "which directory is this component's," and only one
  of them had been moved.
  The same check is deliberately *not* asked of a `Server`: a service
  never has an APK to find under its own directory by definition, so
  asking would report every server "never built" forever. Guarded on
  `matches!(component, Component::Apk { .. })` for that reason.
- **The self entry's project is the working directory itself**, so this
  server has to be started from the root of its own checkout -- which is
  where everything else here is driven from, and what the service unit
  sets (`WorkingDirectory=$REPO_ROOT`, not `server/`). Its components then
  say where they live from there, exactly as any other project's do; there
  is nothing special about this one's layout, which is the point of it not
  being `app/` any more. It used to be
  `CARGO_MANIFEST_DIR`, fixed when the binary was *compiled*, so a
  re-clone or a move left the card watching a directory nobody pulls. The
  symptom is why this is worth remembering: the entry keeps working and
  keeps serving an APK, but silently loses its branch line, its Pull
  button and its commit count, so it reads as an app that simply never has
  an update. It took a "why can't my server see the new commits?" to find.
  Startup warns when there is no checkout at the resolved path, which is
  the only cheap moment to notice.
- **A configured project pointing at this server's own project is a second
  card for it.** `add_app` refuses it (the self entry is in the list it
  checks against, and `self_project` is canonicalized so a symlink can't
  make one directory look like two), so this only survives from a config
  written while the self entry resolved somewhere else. It is not merely
  redundant: it is not the self entry, so `is_self` is false and a build
  through it would restart this server partway through its own build. `build_entries` warns rather than dropping
  it -- dropping it would leave a config entry with no card and so no way
  to remove it from the phone. The built-in card has no Remove button, so
  the duplicate is the one that does.
- **A project says where its own state is; nothing infers it.** The
  `resources:` declaration points at a RON file, a script that prints the
  same RON, or an inline struct, and it holds what a project would
  otherwise write down in several places -- its `name`, and its `data` and
  `config` directories when those are not `$XDG_*_HOME/<name>`. **The file
  is the project's, not this server's**: its own code is meant to read the
  same one, which is why `ResourceFacts` ignores unknown keys where
  `Declaration` sets `deny_unknown_fields`, and why it goes through
  `wg_app_link::format` like every other RON here.
  A project that says nothing gets nothing -- the dialog reports not
  knowing rather than filling in a directory. Two guesses were tried and
  both rejected: the config key (`updater`, `app`, which resolve to
  `~/.config/updater` and `~/.local/share/app`, neither of which exists)
  and the checkout's directory name (right for both projects here, and
  still an inference presented as a fact). Iris's call, and the reason is
  the failure mode: a wrong path does not error, it reads as "this
  component keeps nothing here".
  Read **off the request path**, through `crate::checks` beside the git
  and service checks, because `Script` spawns a process and `/manifest` is
  fetched on every open, resume and Refresh. It is in `checksPending` with
  the other two -- counting only some of them is the bug that once left
  components with no buttons after a restart. And it is part of the
  acceptance gate, since a `Script` runs: comparing the whole declaration
  rather than the variant means changing which file is read is a change
  somebody is asked about.

- **Uninstall can take three things away, and the path on screen is the
  only guard.** The dialog offers logs, data and config as separate
  toggles -- only logs on by default, and ticking data forces logs on,
  since a service that writes its log inside its own data directory would
  lose it either way. `purge.rs` is the whole of it. Where the data and
  config are comes from the project's `resources:` declaration, above; a
  project that does not say gets two disabled toggles saying so, and the
  four reasons a toggle is disabled are distinguishable on screen because
  "couldn't read the resources" is a fault to fix while "doesn't say" is
  an ordinary project.
  A path is removed **wherever it points**, with no check that it
  sits under the XDG directories -- Iris's call, over the alternative of
  refusing anything outside them. What replaces that check is the dialog
  showing each resolved path before the button can be pressed, so the
  phone never asks for a path nobody saw. Anything that stops the path
  being displayed removes the only guard there is.
  Log paths are collected *before* the uninstall runs, because the script
  that reports where a service's log lives is the thing being removed.

- **The self entry can't be removable.** The app can only be updated
  through this server, so an updater that can drop itself from its own list
  strands the installed copy (recovery is a manual reinstall over the
  bootstrap port).
- **`/self` and `/self/apk` are a frozen contract, and the only rescue the
  app has.** Every other route is reachable only by an app new enough to
  understand it: change what `/manifest` says and an older app cannot read
  the list, which is where the button that would replace it lives. So the
  path that fetches a newer app depends on nothing likely to change --
  two numbers, no nesting, no variants, no query parameters -- and the app
  asks it on every launch and again after building its own project, which
  is the moment the newer copy exists and the server it must keep talking
  to has just changed. Add fields at your peril and never rename one;
  anything richer belongs on a route an old app never calls. It does not
  survive a changed CA, port or token, since those break the connection
  before any route is reached: those stay one-way doors and `--download`
  stays their answer.
  **It is also how the built-in card downloads its own APK**, not just
  how the rescue dialog does. `downloadApk` sends the built-in project to
  `/self/apk` rather than to the manifest's route, because this is the
  one download whose server changed underneath it between the build and
  the request -- so it takes the route nothing can rename. It carries no
  `component` or `variant` and does not need to: that project has one
  APK, and which build of it to serve is the mode set on the build
  machine.
  **And the card's Update installs this app itself**, rather than
  building it and leaving a dialog to offer it. Which means it waits for
  the server *twice* (`awaitServerBack`): once before fetching, and again
  before handing anything to the installer. The restart is deferred a
  couple of seconds past the build and then held off for as long as an
  APK is going down the wire, so it lands either just before the download
  or in the gap between the last byte and the install -- a download that
  dies partway reads as the update having failed at the moment it worked,
  and an install offered mid-restart replaces this app while the server
  it must talk to is down. Bounded, and it goes ahead anyway when the
  wait runs out: by the second one the APK is already here, and never
  installing it would be the worse failure.
  **The dialog is now offered only when the built-in card cannot offer
  it** -- that card missing from the list, or no list at all because this
  app is too old to read what the server now sends. `AppListScreen`
  derives that from `manifestState` and reports it up; `Loading` counts
  as present, since not having found out yet is not an answer and a
  dialog flashing up during every load is exactly the noise it exists to
  avoid. Iris's call, 2026-09-02: the card's own Update should do the
  work, "though keep the modal if the dev updater card itself is missing
  or fails to load."
- **There is deliberately no general migration mechanism.** Iris's call:
  each app decides its own, and the updater's job is only to be able to
  get both halves of *itself* to the next version. A project that renames
  or restructures its service is uninstalled and reinstalled from the card,
  which works because dev-updater is not the thing being stopped.
- **The bootstrap link (`--download`) serves the APK already on disk and
  builds nothing.** Shipping a stale one there is the expensive mistake,
  because a fresh install is not enrolled and everything that would replace
  it -- Update, and the QR scan that enrols a device at all -- lives in the
  copy being installed, so the only way out is another trip through
  `--download`. It cost a round of "the scanner fix never landed" when the
  fix was in the checkout the whole time. The startup line names the
  variant and the build's age for that reason; run `./app/build-apk.sh`
  first.
- **Changing the pinned CA is a one-way door for any installed copy.** Same
  reason. `wg_app_link::certs` deliberately won't regenerate an existing
  CA; only the leaf is reissued, on every start.
- **State is per machine, not in the repo**:
  `$XDG_CONFIG_HOME/dev-updater/{config.ron,certs}`, owner-only. This
  repo is shared with a VM at a different path, so a shared config hands
  each machine the other's project paths, and a CA key in it is one that VM
  could sign with. The APK pins the CA of whatever machine builds it
  (`app/build-apk.sh`, `DEV_UPDATER_CA` overrides).
- Freshness is the APK's **mtime** versus `PackageInfo.lastUpdateTime`, not
  a version code — these are ad hoc rebuilds with nothing bumping a
  version. It follows that a build and an install landing in the same
  second can briefly read as "update available"; that's inherent, not a
  bug to chase.

## Running the server for real

`./start.sh` is the shortest way back to a working state: it builds both
halves, rewrites the service unit (so a pull that changes the service script
takes effect) and starts or restarts it. Use it after a pull, or when the
running server and the checkout have drifted.

To run one by hand instead, launch it **from the repo root** -- that is
where it looks for its own app project (`app/`), and started anywhere else
its own card has no checkout -- and **fully detached from the calling
shell**, not via an agent's background-task tracking, which ties its
lifetime to the session:

```sh
setsid nohup /path/to/dev-updater </dev/null >server/dev-updater.log 2>&1 & disown -h
```

Verify with `ps -o pid,ppid,pgid,sid,tty,comm -p <pid>` — detached means
`PPID 1`, its own `SID`, `TT ?`. Check for a stale instance first
(`pgrep -af "[d]ev-updater" || true`): a second instance fails to bind and
silently leaves the old one answering. Stop with
`pkill -f "[d]ev-updater" || true`.

**Run `pgrep -f`/`pkill -f` as a command of its own, with nothing else on
the line, and bracket the first character of the pattern**
(`"[d]ev-updater"`). Both also exit nonzero when nothing matches, which is
the normal case; `|| true` that.

Why the pattern alone isn't enough, since this keeps catching people: each
command runs as `bash -c '<the whole command text>'`, so the wrapper's own
argv contains every word you wrote. An unbracketed pattern therefore
matches the shell running it — `pgrep` reports phantom matches and `pkill`
kills that shell outright (exit 144, truncated output, reads exactly like
the thing under test having crashed). Bracketing the pattern fixes only
*that* word: any **other** mention of the same string on the line re-arms
it. And you usually can't fix the other mention, because it is the real
command —

```sh
# Kills its own shell: the pattern is bracketed, but the path below
# still puts the plain string "target/debug/dev-updater" in the argv,
# and bracketing *that* would change which file gets executed.
pkill -f "[t]arget/debug/dev-updater" || true; setsid nohup ./target/debug/dev-updater …
```

so the only rule that always works is to keep them in separate commands:
one to stop, one to start.

One more way this bites, which bracketing does nothing about: `-f` matches
the **whole command line**, so it also catches processes that merely have
the string somewhere in a path. An agent's scratchpad directory is named
after the repository, so anything launched with a path through it -- the
Android emulator, say -- matches `[d]ev-updater` and gets killed along
with the server. Match the binary rather than the project when stopping
one: `pkill -f "[t]arget/debug/dev-updater"`.

## Environment notes

The machine itself — where the Android SDK is, that each command runs in a
fresh shell so exports have to be chained, and each repo having its own
emulator — is described once in `~/.claude/MACHINE.md`, which every
session reads. What follows is what that means here.

- `app/android-env.sh` applies the SDK override unconditionally, and
  `.claude/settings.json` puts `platform-tools` alone on `PATH` so `adb`
  works without sourcing anything. Everything else needs the env script.
- **This repo has its own AVD, named `dev-updater`**, which is what
  `app/run-android.sh` and `app/enroll-emulator.sh` default to; every
  Android project on this machine has one named after it, and none of them
  share. Two sessions in two repos on one emulator each silently replace
  the app the other just installed, which reads as a build that never
  landed rather than as interference.
- **To exercise installing, add a throwaway test app**, not another repo's
  real one. This server's whole job is putting *other* projects' builds on a
  device, so trying that out needs something to install -- and a scratch
  project you wrote is better for it than a real app, because you can make
  it do whatever the case under test needs: fail its build, produce two
  variants, change its label, bump nothing at all. A real app only does what
  it happens to do. It goes in the list like any other project, so nothing
  here needs to know it is a test.
- **To actually drive the app on the emulator, use
  `app/enroll-emulator.sh`** rather than working enrolment out again. There
  is no camera to scan the QR with, so it fires the `devupdater://enroll`
  intent the app already accepts, generating a token once and adding it to
  `config.ron` beside any real phone's. Two things it exists to stop you
  rediscovering: `adb shell am start -d` loses everything after the first
  unescaped `&`, because the URI reaches the *device's* shell; and the
  server must be started with `--bind 0.0.0.0` for the emulator to reach
  it at all, since it otherwise binds wg0 and the emulator has no route
  there. The script's header has the detail.
  **It names the device rather than assuming there is only one**, which
  matters more now that every repo has its own AVD: a second emulator
  running beside this one makes every bare `adb` call exit 1 with "more
  than one device/emulator", and the version that ran `adb get-state` read
  that as *no emulator running* -- the opposite of what had happened, which
  sends you off to start a third. It resolves the serial by AVD name the
  way `run-android.sh` always has, defaults to `dev-updater`, takes
  `--avd NAME` or `AVD_NAME=`, and when it cannot find that one it lists
  what *is* attached instead of guessing which of the three situations it
  is in.
- The server needs `aapt2` (SDK build-tools) to add an app, and, for an
  app that needs stripping, `llvm-strip` (NDK), `zipalign` and
  `apksigner` (build-tools) **and `~/.android/debug.keystore`** -- the
  last is the least guessable of the four, since it belongs to no SDK,
  is created as a side effect of any Gradle Android build, and is
  consumed by the pipeline's final step. A machine that builds APKs and
  has no keystore is odd rather than new: something removed it after the
  build, and the APK it produced is signed by a key that is now gone. **It has to find
  both without an environment**, because the way this server usually
  starts is from a service manager, and one hands its process a scrubbed
  environment: measured in the Gentoo guest, an OpenRC user service gets
  19 variables with neither `$ANDROID_HOME` nor `$ANDROID_NDK_HOME` among
  them, and a systemd user unit inherits an equally bare one unless
  somebody imported theirs. Neither branch of `service-default` passes any
  through, so this is not a Gentoo quirk and "export it in the unit" is
  not the fix. `HOME` does survive, which is what the home-relative
  candidates in `sdk.rs` are for. The trap it produces is a build that
  *succeeds* and a download that 500s -- a project's own build script
  sources its `android-env.sh` and repairs the environment inside its own
  process, while the strip runs in-process here and sources nothing.
  So `sdk_roots` lists candidates and every lookup takes the first that
  actually contains the tools it needs: an `$ANDROID_HOME` that exists is
  not one with the tools in it (this machine's system-wide
  `/opt/android-sdk` has a `build-tools/36.0.0` holding nothing but
  `package.xml`), and taking it blamed `zipalign` for a root that should
  never have been chosen. Filter *before* taking the newest, too --
  `newest_child_where` exists because filtering the single newest child
  can only reject that one directory, so `Sdk` sitting beside
  `android-ndk-r27c` in `~/Android` hid the NDK from a message that said
  "install one".