fix(sight): avoid false interruption signals

[haoowa]

Description

Avoid false interruption records when a successful assistant response mentions troubleshooting words such as timeout, auth, or rate limit. The detector now only scans structured provider error text and non-2xx response bodies for runtime error classifications.

This PR is split out of #1349 so the interruption classifier fix can be reviewed independently. It also 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 interruption::detector --locked --quiet
• cargo clippy -q -p agentsight --all-targets --locked -- -D warnings

Additional Notes

Split from #1349 following review feedback to keep capture fixes and the grader feature reviewable in smaller units.

[jfeng18]

Thanks for splitting this out — the false-positive suppression goal makes sense. One concern on the recall side:

After this change combined_error only folds in the response body when status_code >= 400 (detector.rs:107-112), and call.error is itself only populated for status_code >= 400 (call_builder.rs:137). So for an HTTP 200 stream that commits the status line and then emits a mid-stream error envelope, both inputs to combined_error are empty — none of the AuthError/RateLimit/ServiceUnavailable/ContextOverflow keyword rules, nor the generic LlmError fallback (detector.rs:284), can fire.

Concretely: Anthropic Messages streaming returns 200 then event: error / {"type":"error","error":{"type":"overloaded_error"}}; OpenAI-compatible proxies stream {"error":{...}} chunks at 200. These land in raw_body and were classified before this PR (e.g. ServiceUnavailable). After it they are either reclassified as SseTruncated (when duration ≥ sse_min_duration_ns) or, for fast-fail (<1s) errors, produce no interruption event at all. Since agents almost always stream, this looks like the common interruption path.

(Reproduced pre/post against the SSE parse path.) Flagging in case the 200-status-with-error-body case is collateral rather than intended.

[jfeng18]

Thanks — this closes the recall regression I raised in round 1. I verified by building and running this time, not just reading the diff:

• cargo test --lib interruption::detector → 42 passed / 0 failed; both new tests (test_detects_anthropic_sse_error_event_in_200_stream, test_detects_structured_sse_error_object_in_200_stream) green.
• Mutation-checked the fix: neutralizing detector.rs stream_error extraction to None flips both tests (detected 0 vs expected 1), so they genuinely gate the fix rather than passing formally.

Also confirmed my main worry does not happen: a >1s 200-SSE error stream is not double-classified — rule 7 (stream_error.is_some()) emits one event and returns before the SseTruncated rule is reached, so exactly one event fires.

Two non-blocking notes for later (not for this PR):

1. Extraction is gated on status_code < 400 && is_sse, so a non-SSE HTTP 200 carrying a buffered (non-streamed) error envelope is still not scanned — same recall class, different transport. Probably a separate follow-up rather than widening this diff.
2. The test_ignores_* false-positive guards all run with is_sse=false, so they don't exercise the new SSE stream_error path; the behavior is correct today but not locked by a committed test.

LGTM on the recall fix.

[jfeng18]

Recall regression from round 1 is closed and verified by build + tests + mutation (42 passed; neutralizing the fix flips the two new SSE tests). No blockers; the two notes above are non-blocking follow-ups. LGTM.