fix(csa-acp,hooks): meaningful-activity timeout + post-merge target cleanup (#854)

[RyderFreeman4Logos]

Summary

Two related changes on this branch:

1. csa-acp — codex initial-response stall detection (Closes #854)

Port initial-response stall classifier to codex-acp transport. Iterated through 4 rounds of review:

• R1 c54e86bf — ported stall detector + retry
• R2 6be39199 — gated classifier on empty pre-timeout stderr (symmetric to gemini gemini ACP initial-stall classifier can overwrite specific stderr failure #912)
• R3 562c2205 — switched watchdog timer to last_activity (respect stderr liveness)
• R4 9c53af31 — introduced separate last_meaningful_activity that excludes protocol-only notifications (AvailableCommandsUpdate etc.) but includes stderr + eligible events

Net: codex ACP stalls are now classifiable and retryable without spurious kills on live-but-slow startups.

2. hooks — post-merge target cleanup

After just install succeeds in scripts/hooks/post-merge-rebuild.sh, clear ./target/ contents via find target/ -mindepth 1 -delete. The ./target symlink itself (pointing to secondary disk on obj host) is preserved; only the resolved contents are removed. Per explicit user directive to reclaim mirror-disk space post-merge.

Review trail

• Final review 01KPNPT8PD41CJ5P8H62Q4SWFY on 9c53af31: decision=fail but findings=[] and severity_counts all zero (known parser FP pattern).
• Prose summary flagged Rust rule 010 "preserve target/" tension with the hook change. Per CLAUDE.md instruction hierarchy, explicit user directive overrides general rule guidance.

Test plan

• cargo test --workspace passed in round-4 commit session
• just pre-commit passed
• Version bumped to 0.1.444

🤖 Generated with Claude Code

[RyderFreeman4Logos]

Merge audit trail — parser FP + user-directive override

Final review session 01KPNPT8PD41CJ5P8H62Q4SWFY on HEAD 9c53af31:

• decision=fail with empty findings=[] and severity_counts all zero → parser FP (known review-verdict parser: decision=fail when reviewer summary explicitly says CLEAN but full.md lacks CSA structured findings section (post-#909) #946 class).
• Prose summary raised concern about Rust rule 010 ("preserve target/") conflicting with the hook's post-install cleanup.

Decision: merge per severity-tiered criteria (findings=0, severity=0).

Rule 010 tension: Per CLAUDE.md instruction hierarchy (User's explicit instructions > Superpowers skills > Default rules), the user's explicit directive to "删除 ./target/* 但不删除软链接" takes priority. Context: user's ./target is a symlink to /ssd/mirror-rootfs/... on a secondary disk; cleanup reclaims mirror-disk space on merge, while the symlink stays intact so cargo build still routes to the secondary disk. The hook also only fires OUTSIDE CSA sessions (CSA_SESSION_ID guard).

Merging with --merge (preserves per-commit audit trail, including the 4 rounds of #854 iteration).

Closes #854.

[gemini-code-assist]

Code Review

This pull request implements a more granular activity tracking mechanism for ACP sessions to support initial response timeouts, particularly for the 'codex' tool. It introduces a last_meaningful_activity timestamp that is updated only by specific agent events (like messages or tool calls), allowing the system to distinguish between protocol-level chatter and actual progress. The changes also include logic in the executor to detect and retry initial stalls for codex. Review feedback correctly identifies code duplication where the event filtering logic was copied into the executor crate; it is recommended to make the original function public and reuse it to maintain a single source of truth.

[gemini-code-assist]

To allow this function to be reused in csa-executor and avoid code duplication, please make it public. You may also need to ensure the client module is public.

Suggested change

	pub(crate) fn event_counts_as_initial_response(event: &SessionEvent) -> bool {
	pub fn event_counts_as_initial_response(event: &SessionEvent) -> bool {

[gemini-code-assist]

This function is a duplicate of event_counts_as_initial_response in csa-acp/src/client.rs. To avoid code duplication and maintain a single source of truth, this duplicated function should be removed.

You'll need to make event_counts_as_initial_response in csa-acp public to be able to use it here.

[gemini-code-assist]

To complete the deduplication, please update this call to use the function from the csa-acp crate.

Suggested change

	.any(event_counts_as_codex_acp_initial_response)
	.any(csa_acp::client::event_counts_as_initial_response)