feat(agent): consume the out-of-band hermes_confab_notice response extension

[Kyzcreig]

What

Hermes-side consumer for the out-of-band confab-catch notice defined in
claude-bpx docs/SPEC-confab-marker-out-of-band.md v1.

A bridge that detects and strips self-fabricated scaffold text from a model reply
must still tell the user — that signal is load-bearing for triage. Today it is
delivered in band, appended to assistant content, which mutates conversation
history and gets replayed upstream on any full-history recovery. This PR moves the
notice to a versioned response-envelope extension: a status event for the current
turn, and presentation-only metadata on the assistant row for historical triage.

Consumer-first is mandatory per the spec — this lands before any producer emits
the field. With today's bridge (no extension field) behavior is unchanged, and
that is pinned by a test.

Transport contract

{"hermes_confab_notice": {"version": 1, "kind": "scaffold_confab_removed",
                          "request_id": "3b264082", "scope": "visible",
                          "grammar": "inbound"}}

Non-stream: top-level on the completion. Stream: on the final usage chunk
(choices: []).

Changes

File	What
agent/confab_notice.py	New. The single validation gate. Fails closed on unknown version, wrong kind, bad scope, non-string/empty/oversized request_id, bad grammar. Returns a fresh dict of only the five contract keys, so a provider cannot smuggle extra fields into display_metadata.
agent/chat_completion_helpers.py	Stream path scans chunks for the extension, accepts at most one per response (a second is logged and dropped), forwards it on the synthetic completion. build_assistant_message stamps display_kind='confab_notice' + versioned display_metadata.
agent/transports/chat_completions.py	Same extraction/validation for non-stream completions, preserved in NormalizedResponse.provider_data.
agent/transports/types.py	NormalizedResponse.confab_notice accessor, alongside the existing codex_* / reasoning_* ones.
agent/conversation_loop.py	Emits the status line once per notice via _emit_status (CLI + TUI + gateway). A request_id ledger stops a retry/fallback re-normalizing the same response from double-emitting.
hermes_cli/cli_agent_setup_mixin.py, apps/desktop/src/types/hermes.ts	Render the new display_kind in the CLI resume recap and the desktop type union.

Assistant content is never touched — no suffix, no prefix.

Replay exclusion

Both presentation fields are already stripped from every outgoing copy in
conversation_loop's api_msg builder. The spec's own line citation for that
strip moved twice, so the new contract test pins it by symbol — an AST walk
for api_msg.pop("display_kind"/"display_metadata") inside the function that also
calls _clone_message_for_send — plus a wire-level e2e asserting the notice never
appears in outgoing messages[].

Per spec §consumer step 4 the flush is not modified; the existing
display_metadata pass-through is asserted instead, so a future refactor that
drops it fails loudly.

Verification

tests/agent/test_confab_notice.py              42 passed
tests/agent/test_confab_notice_e2e.py           8 passed  (stream + non-stream)
tests/agent/test_confab_notice_replay_strip.py  5 passed
tests/tui_gateway/test_confab_notice_render.py  5 passed

The e2e suite drives the real agent loop against an in-process mock provider
returning the v1 object in both wire shapes, and asserts on the captured
outgoing request bytes, not on internal state.

Mutation proofs (the tests are not vacuous):

1. Delete the two api_msg.pop lines → both AST pins and both wire e2e tests go
   red (5 failed, 8 passed). Restored.
2. Delete "display_metadata" from the flush row dict → the persistence pin and
   both persistence e2e tests go red (3 failed, 6 passed). Restored.

Neighbor suites, all green:

tests/agent/transports + sidecar     18 files, 420 passed
streaming + display                   9 files, 127 passed
CLI resume + gateway history          6 files, 244 passed

Not in this PR

The bpx producer. It is a separate card, blocked on this one — the bridge must not
emit the field until this consumer is deployed.

[Kyzcreig]

FleetReview

Confidence: 3/5

Findings

• P1 agent/chat_completion_helpers.py:5225 — Duplicate Accepted
• P1 agent/conversation_loop.py:7670 — Alert Suppression
• P1 hermes_cli/cli_agent_setup_mixin.py:801 — False Notice
• P1 apps/desktop/src/types/hermes.ts:586 — Desktop clients never surface the confabulation notice
• P1 tests/agent/test_confab_notice.py:201 — Provenance Untested
• P2 tests/agent/test_confab_notice_e2e.py:243 — Missing Marker
• P2 tests/agent/test_confab_notice_e2e.py:269 — Incomplete Validation
• P2 tests/agent/test_confab_notice_replay_strip.py:145 — Self-Testing
• P1 tests/agent/test_confab_notice.py:137 — Streaming Untested
• P2 tests/agent/test_confab_notice_e2e.py:188 — Mode Coverage
• P3 agent/confab_notice.py:18 — Module docstring's fail-closed list claims duplicate-notice behaviour the code does not implement
• P2 tests/agent/test_confab_notice_e2e.py:119 — Fixture never calls server_close(), and server/tempdir setup sits outside the try/finally
• P1 agent/chat_completion_helpers.py:2488 — Persisted notices are invisible in reloaded TUI and desktop sessions

---

FleetReview provenance · models: B=gpt-5.6-sol, C=claude-code-opus-5, F=gpt-5.6-sol, G=grok-4.6 · cost: $35.76 · duration: 1h 12m 29s · rounds: 2 · files examined: 11