fix(sight): persist idle streams and tool results

[haoowa]

Description

Persist idle in-flight streamed calls so manually interrupted conversations remain visible as pending evidence instead of being silently dropped. This also preserves Anthropic tool_result blocks as structured input message parts so downstream scoring can distinguish actual tool failures from historical raw context.

This PR also adds regression coverage for oversized request-body eviction and includes the current-toolchain mechanical AgentSight clippy fixes required for this split PR to pass -D warnings independently.

Related Issue

Refs #1304

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Refactoring (no functional change)
• Performance improvement
• CI/CD or build changes

Scope

• cosh (copilot-shell)
• cosh-ng (cosh-ng)
• sec-core (agent-sec-core)
• skill (os-skills)
• sight (agentsight)
• tokenless (tokenless)
• ckpt (ws-ckpt)
• memory (agent-memory)
• anolisa (anolisa-cli)
• skillfs (SkillFS)
• Multiple / Project-wide

Checklist

• I have read the Contributing Guide
• My code follows the project's code style
• I have added tests that prove my fix is effective or that my feature works
• I have updated the documentation accordingly
• For cosh: Lint passes, type check passes, and tests pass
• For cosh-ng: cargo clippy --all-targets -- -D warnings and cargo fmt --check pass
• For sec-core (Rust): cargo clippy -- -D warnings and cargo fmt --check pass
• For sec-core (Python): Ruff format and pytest pass
• For skill: Skill directory structure is valid and shell scripts pass syntax check
• For sight: cargo clippy -- -D warnings and cargo fmt --check pass
• For tokenless: cargo clippy -- -D warnings and cargo fmt --check pass
• For memory (Linux only): cargo clippy --all-targets -- -D warnings, cargo fmt --check, and cargo test pass
• For anolisa: cargo clippy --all-targets --locked -- -D warnings, cargo fmt --all --check, and cargo test --locked pass
• For skillfs: cargo fmt --all --check, cargo clippy --workspace --all-targets -- -D warnings, and cargo test --workspace pass
• Lock files are up to date (package-lock.json / Cargo.lock)

Testing

Ran on the Linux AgentSight test host from an archive of this split branch:

• cargo fmt --all -- --check
• cargo test -p agentsight aggregator --locked --quiet
• cargo test -p agentsight genai::call_builder --locked --quiet
• cargo clippy -q -p agentsight --all-targets --locked -- -D warnings

Additional Notes

Split from #1349 following review feedback to keep capture fixes separate from the grader API and dashboard work.

[jfeng18]

Re-checked this one closely since it's timing-bounded — it holds, with a scoping note.

New in this PR: drain_and_persist_idle_connections (unified.rs:1287) persists a fresh-generate_id() pending row for a live process once a stream idles past idle_timeout (60s, config.rs:61), and pops the connection from the aggregator. Since complete_pending keys on the real call_id, that orphan row is never resolved even if the stream resumes and completes successfully.

Consequence: ProcMonEvent::Exit carries no exit code (procmon.rs:47-53), so handle_agent_crash_detection (unified.rs:861) fires on every tracked-agent exit and sweeps status='pending' rows for the pid into InterruptionType::AgentCrash (Critical, types.rs:77), timestamped at exit. So a stream that idled >60s, resumed and completed fine, followed by a clean /exit within ~[t0+60s, t0+300s], yields a spurious Critical agent_crash on a successful session. After ~t0+300s the stale scanner flips the row to interrupted and it no longer trips the sweep, so the window is ~4 min.

Attribution, to be fair: the exit-without-exit-code gate and "pending row at clean exit → agent_crash" behavior already exist on main; this PR doesn't introduce those. What's new is a durable, unresolvable pending row for a successful idled-then-resumed stream, which previously left no row at all. Flagging since a false Critical on a clean exit is costly for a detection product; the deeper "Exit has no exit code" limitation looks pre-existing and probably a separate discussion.

[jfeng18]

The idle-drain fix closes the finding I raised — a clean process exit no longer sweeps idle-drained rows into a spurious Critical agent_crash. I built and ran it rather than just reading:

• cargo test --lib storage::sqlite::genai → 51 passed / 0 failed; test_crash_sweep_ignores_idle_drain_pending, test_list_pending_for_pid(s), and test_complete_pending_promotes_idle_snapshot_by_match_key all green.
• Mutation-checked the gate: removing the AND pending_origin != 'idle_drain' filter from list_pending_for_pid flips test_crash_sweep_ignores_idle_drain_pending (2 rows vs expected 1), so the exclusion is genuinely what prevents the false crash.

Both halves of the original concern are addressed: the PendingOrigin::IdleDrain tag + its exclusion from the crash sweep stops the spurious agent_crash, and pending_match_key + the non-destructive snapshot let a resumed stream reconcile the row instead of it being an unresolvable orphan. Nice.

One thing blocking merge, though — CI is red on Commit Message Lint (not the tests; Test agentsight passes). From the job log:

✖  body's lines must not be longer than 100 characters [body-max-line-length]
✖  found 1 problems, 0 warnings

Both commits trip it — the bodies have lines of 146/359 chars (preserve idle stream snapshots) and 120 chars (persist idle streams and tool results). Rewrapping the two commit bodies at ≤100 chars and force-pushing should turn it green; happy to approve once CI is clean.

[haoowa]

Additional note on the code changes made for the earlier #1394 review feedback.

The review concern was that persisting idle streams could make an active stream
look like a crash-drained pending call, and that a resumed stream might not be
able to reconnect to the idle snapshot later.

The PR now addresses that by:

• Marking idle-timeout persisted rows explicitly as idle_drain via
  pending_origin, separate from normal request capture and dead-PID drains.
• Adding a deterministic pending_match_key so a later completed stream can
  promote the previously persisted idle snapshot instead of leaving duplicate
  or stale pending evidence.
• Changing idle timeout handling from destructive drain to non-destructive
  snapshot, so SseActive state remains in memory and the stream can continue
  if more SSE chunks arrive.
• Excluding idle_drain rows from crash/dead-PID pending sweeps, so a live but
  idle stream is not reclassified as an agent-crash artifact.
• Adding SQLite schema migrations and regression tests for:
  • persisted idle origin and match key
  • promotion by match key
  • crash sweep ignoring idle snapshots
  • resumed SSE streams after idle snapshot

Validation was run on the Linux test host:

• cargo fmt --all -- --check
• cargo test --quiet
• cargo clippy --all-targets -- -D warnings

The latest force-push only wrapped commit message bodies for commitlint; it did
not change the code tree from the reviewed fix.

[jfeng18]

Idle-drain finding closed and verified (build + 51 passed + mutation on the filter). CI now green after the commit-body rewrap; code is byte-identical to the version I verified. LGTM.