modem-stack: Phase-B adoption docs sweep + POLICY amendment + v1.0.0 version bump

[andrescera]

What

Closes out modem-stack's Wave 0+1 (todos 1-10, already merged onto this branch) with the integration housekeeping: doc sweep, the POLICY.md Phase-A→Phase-B scope amendment, and the 1.0.0 version bump for both workspace packages. No source/behavior changes — this PR is docs + one version-bump commit.

• AGENTS.md / README.md: document the uhubctl PowerHook (recovery-ladder rung 4, control/src/backend/uhubctl-power-hook.ts), the hil-cycle HIL harness (bench runbook RB-10), the full RB-9..RB-17 bench-runbook ladder, and add the docs/ESIM-DECISION.md pointer (eSIM is investigate-only, implementation deferred).
• POLICY.md §4: amended from a blanket "Phase A only, no CeraUI/image/apt integration" statement to an adoption-versioned one — integration is authorized starting at the v1.0.0 tag, not before, while §1-3 (no-fork gate, upstream-first) remain permanently binding. The historical fact that every release through v0.2.0 was Phase-A-only is preserved, not scrubbed. README.md and packaging/README.md carry the same amendment so no stale "Phase A: no integration" claim survives outside POLICY.md's own record of it.
• control/package.json (@ceralive/modem-control) and cli/package.json (modem-control-cli): 0.2.0 → 1.0.0, the version POLICY.md §4 and docs/VERSIONING.md reserve for Phase-B adoption.

No tag is cut in this PR. v1.0.0 is released later (a separate workflow_dispatch on release.yml), once CeraUI's and image-building-pipeline's own Phase-B waves are also ready to consume it.

Why

This is the last modem-stack todo in the current effort. The repo's docs/VERSIONING.md already reserved 1.0.0 for "Phase-B adoption"; this PR is what actually enacts that reservation in the version files and in the policy text that currently forbids any downstream integration outright. Bumping now (rather than at tag time) means the tag, when cut, releases exactly what this PR's gates already proved green — no second PR, no last-minute version-file diff at release time.

How to verify

• Pre-check (done before touching any version file): read .github/workflows/{ci-bun.yml,ci-packaging.yml,release.yml} and packaging/ci/*.sh. Confirmed no lane asserts control/package.json's version against an existing tag/npm/deb state on a non-tag (PR/push) build — ci-bun.yml only lints/typechecks/tests; ci-packaging.yml's contract.sh never reads control/package.json (only the four packaging debian/changelog files); the one version-vs-tag assertion (release.yml's publish-npm job) only runs on manual workflow_dispatch with an explicit tag input, unreachable from this PR.
• bun install --frozen-lockfile → exit 0.
• git diff --exit-code -- bun.lock → exit 0 (zero drift; workspace members are referenced by name, not version, in the lockfile's resolution table).
• bunx tsc --noEmit → exit 0.
• bunx biome check . → exit 0 (171 files).
• dbus-run-session -- bun test → 396 pass / 0 fail across 50 files.
• bash packaging/ci/contract.sh → PASS (scaffold, tag-guard contract, dpkg tilde-ordering proofs all green; dch absent in this sandbox, which the script itself detects and documents rather than failing).
• Stale-version sweep: no surviving bare 0.2.0 self-reference in cli/. The 0.2.0 mentions remaining in docs/BENCH.md/docs/VERSIONING.md are accurate as-is — v0.2.0 really is the latest released tag until v1.0.0 is actually cut (a later step), so those are correctly-retained current-state text, not stale claims.

Risks

• Not run in this PR: packaging/ci/{build-bookworm.sh,test-package-contract.sh,daemon-smoke.sh} — these build real .debs in a debian:bookworm container and need a working Docker daemon; this sandbox's Docker context is Docker Desktop, which this stack's own tooling refuses as a runner daemon (same known limitation recorded for image-building-pipeline). This is not a gap in this PR's own gate: those three scripts are release.yml-only (tag-triggered), not part of ci-packaging.yml's PR lane, so nothing in this diff exercises them.
• bun.lock has no diff to commit — bumping a workspace member's own version doesn't touch its resolution-table entry (verified, not assumed).
• The POLICY.md amendment lifts a standing prohibition but authorizes nothing on its own: every actual integration (CeraUI, image-building-pipeline, apt-worker) still needs its own reviewed PR in the receiving repo, gated on this repo's v1.0.0 tag existing.

---

Update — three further commits on this branch

The branch advanced past the version bump before merge. The PR now carries real source
changes in addition to the docs/policy/version work described above, so the "docs +
one version-bump commit" framing at the top applies only to the original three commits.

• 9b9184b feat(usage): setUsagePolicy write surface — the WRITE half of the
  data-usage surface (control/src/backend/usage/policy-write.ts + policy-store.ts).
  It persists a slot's cycle day and advisory threshold to a local 0600 policy file
  and, when a live UsageSampler is supplied, applies them in the same call. It writes a
  local file and never the modem, because ModemManager exposes no data-usage API at all:
  verified against a live MM 1.24.2 that the only Setup/threshold surface on the object
  is Modem.Signal.Setup (rssi-threshold / error-rate-threshold — radio quality, not
  bytes), and the only byte counters are the per-bearer read-only Stats, which reset on
  every connection and so cannot carry a monthly cycle. Typed results rather than throws
  (called from an RPC boundary), tri-state fields (undefined leaves alone, null
  clears), order is load → validate → persist → apply, and an applied policy outranks a
  later observation carrying the old one for the process lifetime.
• 4d53b09, eb71039 docs(fm350) — record the real FM350-GL observed on USB
  (Citation 6) and close the adapter-mediated decision stop in docs/FM350-DECISION.md.
  Documentation only; the modem stays documented-deferred, not supported.

Verification on the merged head (eb71039)

• bun install --frozen-lockfile → exit 0, bun run lint (Biome, 175 files) → exit 0,
  bun run typecheck (tsc --noEmit, strict + exactOptionalPropertyTypes) → exit 0.
• bun test → 448 pass / 0 fail, 9274 expect() calls across 52 files.
• Both PR checks green pinned to this SHA: Lint, typecheck, test (control + cli) and
  Packaging contract (bookworm).
• npm pack ./control --dry-run → @ceralive/modem-control@1.0.0, 121 files — the
  exact tarball CeraUI's consumer-side skew test was proven against before publish.

The .deb-building lanes remain release.yml-only (tag-triggered) and are unchanged by
these commits; they run for real when v1.0.0 is dispatched.