mobile: consume agent mention acks (kind 44102)

[mfethe1]

Agents publish NIP-MR mention acknowledgements (kind:44102) when they accept or decline an @-mention. Desktop consumes them; mobile had zero references to 44102, so a phone user could not tell a decline from a delay. This adds mobile consumption.

EventKind.agentMentionAck is added to the subscribed channel kinds so acks reach the client, and to channelAuxEventKinds so 44102 overlays rather than rendering as a timeline row. Acks are routed in the live handler ahead of the channel-visibility and unread work, mirroring desktop's ordering, so an ack can never bump an unread badge, reorder channels, or change the last-message preview. A new community-scoped in-memory store keys outcomes by mention event id then signer pubkey, making double delivery idempotent by data shape.

Deliberate scope cut vs desktop: no silent verdict and no client timer. Mobile suspends often, so a timer would fire false "nobody picked this up" verdicts after resume. Only received facts are stored, so a dropped ack degrades to "no answer yet" — never to a false decline.

No migration, no schema, no protocol change. The ACP publisher and desktop are untouched.

Security invariants

• Attribution is to the event signer, never the p tag (which carries the mention's author and is forgeable as an identity claim).
• Only acks whose signer was actually tagged in the mention are applied — the relay is pure fan-out and cannot check agent-ness.
• Only the local identity's own messages surface an outcome.
• reason is untrusted: clamped to 200 chars at the parse boundary, rendered as plain Text (no markdown, no link handling).
• Unknown status is ignored, never coerced to declined.
• Store is bounded at 512 event ids with oldest-first eviction.

Verification

• flutter test (full mobile suite): 1758 passing, 14 failing.
• The same 14 failures occur on untouched upstream main (1647 passing, 14 failing). Failure name sets are identical — comm -23 ours baseline is empty. No regression. Pre-existing failures live in channel_detail_page_test.dart, channels_page_test.dart, search_page_test.dart, remind_me_later_sheet_test.dart.
• flutter analyze: 6 issues, unchanged from baseline (4 × deprecated_member_use, 1 × valid_regexps, 2 × unawaited_return_in_try_block) — all pre-existing, none in changed files.
• cargo test -p buzz-acp --lib pool::tests: 131 passed, 0 failed. Publisher tag shape unchanged.
• New store test: 19 cases covering signer attribution, untagged-signer rejection, unknown/absent/empty/wrong-case status, reason clamping, idempotent double delivery, accept-outranks-decline, bounded eviction.
• Widget render verified via an ephemeral golden capture of both accepted and declined states (indicator found, non-blank raster). The capture harness was not committed.

Not verified

• No emulator/simulator run and no live relay test. No real 44102 event from a real agent was observed end to end; acks are synthesized in tests from the publisher's tag shape.
• No performance profiling of the 512-entry bound.

Known limitation

An ack landing while its channel is not in the live set is dropped. This is strictly better than today, where every ack is dropped, and it shows as "no answer yet" rather than a false decline.