Bound and clear the ACP compaction prose buffer

[bradhallett]

AGENT GENERATED

Human comments

Follow-up to #2293, addressing the two review findings SlopCop left on the merged commit (eece92f8):

1. Major — "OMP text rule applies to all ACP agents": already fixed on main by Scope compaction outcome parsing to OMP #2957, which moved end-turn compaction classification behind the dialect (compactionOutcomeForEndTurn(dialect, …) with dialect.id !== "omp" short-circuiting to completed). No further change needed; this PR builds on that shape.
2. Medium — "unbounded compaction message text, no cleanup in finishCompaction": still open on main; this PR fixes it.

What was wrong

• handleAgentNotification accumulated every agent_message_chunk streamed during a compaction prompt into session.compactionAgentMessage for every dialect, even though only the omp dialect ever reads the text — unbounded growth on a long-lived session.
• The accumulation had no size limit.
• The buffer was only reset at startCompaction, so after a compaction completed, the text stayed resident until the next compaction began.

What changed

• Accumulation is now gated on collectsCompactionAgentMessage(session.dialect) — the dialect module owns the "does this dialect parse compaction prose" decision, so non-omp dialects carry no classifier state at all.
• Accumulation clamps at 64 KiB (COMPACTION_AGENT_MESSAGE_CAP), keeping the head of the message: omp's failure/no-op prose ("Compaction failed: …", "Nothing to compact") leads its compaction report.
• The buffer is cleared in finishCompaction — the single settlement funnel that success, failure, and interrupt all route through. The now-redundant reset at startCompaction is removed; session-creation initialization covers the first compaction.
• fake-acp-agent.mjs gains FAKE_ACP_COMPACT_AGENT_MESSAGES (JSON string array; one message per /compact request, last entry repeats) so one session can exercise sequential compactions with different prose.
• SDK bump 0.4.40 → 0.4.41 (bridge dist output changes).

How you verified

• Three new regression tests, each mutation-checked (revert the guard, confirm the test fails, restore):
  • failure prose within the cap still classifies the turn failed — kills a tail-keeping clamp;
  • failure prose past the cap is ignored and the turn completes with thread/compacted — kills cap removal;
  • a second compaction in the same session classifies independently after a failed first one — kills removal of the finishCompaction clear.
• The accumulation dialect gate is behaviorally invisible by construction (no non-omp code reads the buffer), so it is covered by Scope compaction outcome parsing to OMP #2957's existing generic/OpenCode completed-outcome regressions rather than a test that could not fail.
• Package gates: turbo run test typecheck --filter=@bb/provider-bridge-acp --force — 18/18 test files, typecheck clean; oxfmt --check clean; oxlint zero new findings (the pre-existing bb(no-comments) hits in the .mjs harness are identical to main: 15 before, 15 after); check-npm-version-guard PASS at 0.4.41.