feat(acp): add opt-in durable session binding store

[KLari1994]

Summary

Adds an opt-in SQLite session-binding and processed-event store to buzz-acp so a stored ACP session can survive harness restart, duplicate relay event IDs produce one reply, and invalid bindings are discarded instead of silently forking history.

Default remains in-memory. Set --session-store PATH or BUZZ_ACP_SESSION_STORE to enable. The store holds IDs and timestamps only.

Closes PAN-3.

Why

Session maps lived only in memory. A restart always called session/new and forked history. Event dedupe was also in-memory, so a re-REQ after restart could prompt twice.

What landed

Branch feat/PAN-3-session-store on pinned main 0720f5380ce8a6c050afac159f8462c06cd51ab5.

Commits (DCO-signed):

• 5b3fdcd5b feat(acp): add durable session binding store
• a18e48d13 fix(acp): scope session bindings per worker and harden restore
• c11e582ec fix(acp): drop bindings on idle model switch and retire sibling keys

HEAD: c11e582ec17293f0036f4363e1b26d2fdde86c71
Diff: 10 files, +2374 / -39, crates/buzz-acp plus mechanical root Cargo.lock.

New:

• crates/buzz-acp/src/session_store.rs
• crates/buzz-acp/src/session_store/sqlite.rs
• crates/buzz-acp/tests/session_store_restart.rs

Changed: config.rs, acp.rs, pool.rs, lib.rs, queue.rs, Cargo.toml.

Behavior:

• Restore uses ACP session/load only when the adapter advertises loadSession. Failure or missing capability deletes the binding and warns, then creates a fresh session.
• Bindings are worker-scoped. A fresh channel create retires sibling-worker keys first.
• Processed events are marked only after a successful turn.
• Store errors never take the harness down.
• No Pantheon names. IDs and timestamps only.

Verification

Independent gate: cargo test -p buzz-acp session_store, cargo test -p buzz-acp pool, cargo test -p buzz-acp, cargo fmt --all --check, cargo clippy -p buzz-acp --all-targets -- -D warnings all exit 0 at 2026-08-24T10:17:02Z.

This is a generic upstreamable seam. Hosts must pass a stable store path; Desktop/gateway launcher wiring is out of scope.

[KLari1994]

Pantheon / PAN-3 — merge assistance request

This PR is the complete implementation for Linear PAN-3 (durable opt-in buzz-acp session binding store).

• Head: c11e582ec17293f0036f4363e1b26d2fdde86c71 (3 DCO-signed commits)
• Base: main @ 0720f5380ce8a6c050afac159f8462c06cd51ab5
• Local independent gate: cargo test -p buzz-acp / fmt / clippy -D warnings all green
• Author association: NONE — KLari1994 has pull-only on block/buzz (push=false), so we cannot merge or satisfy admin branch-protection ourselves
• mergeable=true, mergeable_state=blocked (branch protection / required checks/reviews)

Could a maintainer please review + merge (squash OK), or advise the preferred contributor path?

Thanks — Syntropic/Pantheon integration is serial-blocked on this landing.

[KLari1994]

Follow-up: CI is waiting on maintainer workflow approval

Re-checked head c11e582ec17293f0036f4363e1b26d2fdde86c71:

Surface	Result
DCO Check	success
CI	completed / action_required — https://github.com/block/buzz/actions/runs/32717764216
Docker image	action_required — https://github.com/block/buzz/actions/runs/32717764208
Mesh Lifecycle	action_required — https://github.com/block/buzz/actions/runs/32717764224
Desktop Release Candidate	action_required — https://github.com/block/buzz/actions/runs/32717764206
mergeable_state	blocked
KLari1994 perms on block/buzz	pull only (push=false)

For first-time contributors from forks, GitHub often holds Actions until a maintainer clicks Approve and run workflows. After CI is green, a maintainer merge (squash OK) unblocks Pantheon PAN-3.

Local verification already green: cargo test -p buzz-acp + fmt + clippy -D warnings.