feat(acp): host-wide session pool with fleet slot cap, idle close and DM auto-publish

[gruming]

Summary

Memory on a host running many buzz-acp harnesses scales with live ACP sessions (for claude-agent-acp, one claude process plus its MCP servers each), not with harness count. This PR adds a connection-pool style cap and idle reclamation so a fleet of 25 harnesses fits on one 32 GB host, plus a DM safety net.

• --fleet-slots N / --fleet-slot-dir (BUZZ_ACP_FLEET_SLOTS, BUZZ_ACP_FLEET_SLOT_DIR): host-wide cap on live sessions shared by every harness. One flock(2)ed file per slot, so the kernel releases a dead harness's slots. Work that finds every slot held waits in the channel queue rather than failing (crates/buzz-acp/src/fleet.rs, new).
• --session-idle-close S (BUZZ_ACP_SESSION_IDLE_CLOSE): close a channel session, releasing its slot, after S seconds without a turn; the next event in that channel starts a fresh session.
• --min-agents M (BUZZ_ACP_MIN_AGENTS): start only M subprocesses eagerly; the rest of --agents spawn on demand when every started agent is busy and channels are waiting, and shut down again after the idle bound.
• --dm-autopublish (BUZZ_ACP_DM_AUTOPUBLISH, default false): in DM channels, when a turn ends without the agent having run buzz messages send, post the agent's trailing reply text as a top-level message so the human never gets silence. Skipped when the agent already posted in that turn (relay lookup, fails closed); channel turns are unaffected.

All new flags — the pool flags and --dm-autopublish — default off, so existing deployments are unchanged. README documents the flags and the fleet example (--agents 8 --min-agents 1 --fleet-slots 20 --session-idle-close 600).

Key decisions / trade-offs

• Slots are flocked files rather than a shared counter so a crashed harness can never leak capacity. This is Unix-only; fleet.rs refuses to start with --fleet-slots > 0 where flock is unavailable.
• Idle close is per channel session, not per agent, so a busy channel keeps its context while quiet ones return their slot.
• DM auto-publish is deliberately fail-closed: any doubt (self-post lookup error/timeout, text empty, agent already sent) means no extra post.

Relation to #6732 (per-thread sessions)

• Rebased onto current main (40220d5). A "live session" is keyed by SessionScope, so under the default --session-policy channel behavior is per channel; under thread policy fleet slots and idle close apply per thread session automatically (a busy thread keeps its context, quiet ones return their slot). On-demand sizing counts distinct pending scopes.
• Nothing upstream was dropped or refactored; conflicts were resolved by re-keying our bookkeeping from channel Uuid to SessionScope.

Deferred / follow-up

• Opening as a draft until the rebased build has run in our fleet for a while. Happy to split the DM auto-publish part into its own PR if preferred.
• --session-idle-close / --fleet-slots help text still says "channel session"; accurate under the default policy, could be softened to "session" if thread policy becomes the default.

Related issue

Addresses the memory growth in #2961 (per-channel sessions never released) and #4577 (subprocesses never reaped) by bounding and reclaiming live sessions host-wide. No existing PR found for a fleet-wide session cap; #7173 and #5982 touch idle/rotation handling but not a shared cap.

Testing

• cargo fmt -p buzz-acp --check, cargo clippy -p buzz-acp --all-targets -- -D warnings: clean.
• cargo test -p buzz-acp: 924 passed, 0 failed (plus 9 integration tests in tests/pool_lifecycle_state.rs; doc-tests: 0), on the rebased commit (new unit tests for fleet slots, min-agents resolution, idle close and dm_autopublish_wanted). Note: CliArgs tests read BUZZ_ACP_* from the environment like the existing ones, so run them with those variables unset.
• Not run: just ci (this host has no GTK/WebKitGTK toolchain for the desktop crates); only the buzz-acp crate was touched.
• Production canary: 25 harnesses on one 32 GB Ubuntu 24.04 host with the flags above since 2026-09-03 (KST), on the pre-rebase build; the rebased build is being canaried next. live sessions capped at 20, idle sessions reclaimed after 10 min, on-demand agents spawn and retire as expected. DM auto-publish verified on one harness for the "agent already posted → skipped, no duplicate" path; the "trailing text posted" path is covered by unit tests only so far.

🤖 Generated with Claude Code

[github-actions]

🔐 Codex Security Review

Status: review required for the current range.

The current range is 4b0744d7f3161d9f045dd67a73e014e4440c2b47...800e694b268500df00b6089e96c066315b28a4b3.
A new review must complete for this exact range. When manual authorization
is required, a Block organization member must comment exactly
@buzz-security-review 800e694b268500df00b6089e96c066315b28a4b3 to authorize a new review.
Any previous review applies only to its recorded range.