refactor(gateway): consolidate interruption-note builders + close WhatsApp reaction gap

[aldoeliacim]

Summary

Phase 0 (of 2) for a reply-gate/smart-reply architecture redesign. This PR is
pure groundwork — zero behavior change by design: refactors, additive
methods, vocabulary consolidation, and new regression tests only. Phase 1
(the actual tool-gated delivery mechanism, config-flagged off by default)
will follow as a separate PR once this lands.

What changed and why

1. Consolidate the interruption/resume system-note construction.
gateway/run.py had three near-duplicate sites constructing the
[System note: The previous turn was interrupted by ...] bracket text — one
of them (the "safety net" fallback) was a byte-for-byte copy of another's
empty-message branch. Extracted into two pure builders
(_build_interruption_system_note, _build_tool_tail_system_note) plus one
dispatch predicate (_resolve_resume_note_kind) that formalizes which of the
three sites should fire. Verified byte-identical output via golden-string
tests captured from the pre-refactor text.

2. Consolidate the silence-marker vocabulary.
cron/scheduler.py's _CRON_SILENCE_TOKENS was an independently-maintained
duplicate of gateway/response_filters.py's LIVE_GATEWAY_SILENT_MARKERS —
same 4 tokens, but two separate declarations that could silently drift (and
per the inline comments, already had once). Cron now imports the canonical
set lazily (avoiding an eager gateway-package import from cron-only paths);
its own positional first/last-line matching logic is unchanged.

3. Close a real gap in the WhatsApp reaction path.
tools/send_message_tool.py already has a generic action='react'/'unreact'
path on main that probes for add_reaction(chat_id, emoji, message_id) /
remove_reaction(chat_id, message_id) on the platform adapter via getattr
— but neither WhatsApp adapter implemented these methods, and
scripts/whatsapp-bridge/bridge.js had no /react endpoint at all. Added:

• add_reaction/remove_reaction on WhatsAppAdapter (Baileys bridge) and
  WhatsAppCloudAdapter (Graph API reaction message type)
• A minimal POST /react endpoint on the bridge (empty emoji retracts)
• latestForChat() on the bridge's bounded message store, so a reaction can
  target "the most recent message in this chat" without an explicit message
  id

4. New cross-mechanism regression tests.
No existing test proved the interruption/resume-note mechanism and the
silence-detection mechanism compose correctly. Added:

• A realistic recovery-report string ("The gateway restarted and is back
  online now...") is never misclassified as an intentional silence marker
• A resumed turn whose model output is exactly a silence token still
  suppresses correctly
• A report that merely mentions a marker word mid-sentence still delivers

How to test

scripts/run_tests.sh tests/gateway/test_resume_note_silence_interaction.py tests/tools/test_send_message_react.py -v
cd scripts/whatsapp-bridge && node --test *.test.mjs

All 37 new/modified Python tests pass; all 24 existing + new JS bridge tests
pass. Full suite run separately confirms no regressions (see verification
note below).

What platforms tested

Linux (WSL2-adjacent dev environment). No platform-specific code paths
touched — the WhatsApp Cloud API reaction method and Baileys bridge endpoint
are both cross-platform Python/Node, no OS-specific syscalls.

Verification note (full suite)

Ran the full test suite via scripts/run_tests.sh both with and without this
diff (via git stash) to confirm no regressions. Baseline on unmodified
main: 32 pre-existing test failures + 9 tests/acp* collection errors
(missing optional acp extra, pip install hermes-agent[acp] not run in
this dev environment) — identical failure set with or without this diff.
ruff check and scripts/check-windows-footguns.py are clean on every
touched file.

Related / overlapping PRs (please read before reviewing)

• feat(whatsapp): add emoji reaction support #38862 (community) adds WhatsApp reactions via a new
  send_message(action='react') schema field + bridge endpoint. This PR is
  narrower and complementary: send_message's react/unreact action already
  exists on main today — this PR only implements the two adapter methods
  that path was already trying to call via getattr and was silently
  failing to find. No schema changes here.
• feat: add WhatsApp quote and status primitives #21977 (feat/whatsapp-status-reply-primitives, our own open PR) has a
  similar add_reaction-style adapter method on a separate, unmerged branch.
  This PR was implemented fresh against current main, not derived from or
  dependent on that branch — no relationship beyond solving an adjacent
  problem.

Happy to help reconcile with whichever of these a maintainer prefers as
canonical if there's overlap concern at review time.