feat(builder-sidecar): log API calls to LOG_DIR for evaluation

[0xdkay]

Summary

Add structured JSONL logging of every builder sidecar API call (run-pov, apply-patch-build, run-test) to OSS_CRS_LOG_DIR for CRS evaluation.

Closes #148

Implementation

• Register /sidecar-logs via libCRS register-log-dir at startup, symlinked to host-mounted OSS_CRS_LOG_DIR
• After each API handler completes, append one JSON line to api-calls.jsonl
• Each entry includes: ts, api, job_id, exit_code, duration_ms, plus api-specific fields
• Handlers return harness/build_id in result dicts (no separate pre-read needed)
• crash uses exit_code > 0 to avoid false positives from handler exceptions (sentinel -1)
• Empty api-calls.jsonl created at startup to distinguish "zero calls" from "sidecar never started"
• Best-effort logging — never blocks or fails the API call

Log entry examples

Bug-finding (crs-bug-finding-claude-code on sanity-mock-c-delta-01):

{"ts":1774296986.79,"api":"run-pov","job_id":"2b9ec6bf26cf","duration_ms":58,"exit_code":1,"harness":"fuzz_parse_buffer_section","build_id":"base","crash":true,"timeout":false}
{"ts":1774297023.24,"api":"run-pov","job_id":"dbe6bf714002","duration_ms":55,"exit_code":1,"harness":"fuzz_parse_buffer_section","build_id":"base","crash":true,"timeout":false}
{"ts":1774297067.71,"api":"run-pov","job_id":"bcee79ed184a","duration_ms":65,"exit_code":77,"harness":"fuzz_parse_buffer_section","build_id":"base","crash":true,"timeout":false}

Bug-fixing (crs-claude-code on sanity-mock-c-delta-01 with POV input):

{"ts":1774300121.36,"api":"build","job_id":"52d872574bec","duration_ms":551,"exit_code":0,"build_success":true}
{"ts":1774300136.70,"api":"run-pov","job_id":"d9f4cce856b6","duration_ms":107,"exit_code":1,"harness":"fuzz_parse_buffer_section","build_id":"base","crash":true,"timeout":false}
{"ts":1774300148.91,"api":"run-pov","job_id":"21fc5831ed93","duration_ms":90,"exit_code":0,"harness":"fuzz_parse_buffer_section","build_id":"52d872574bec","crash":false,"timeout":false}
{"ts":1774300159.91,"api":"run-test","job_id":"bd45c2ec677e","duration_ms":176,"exit_code":0,"build_id":"52d872574bec","test_passed":true,"skipped":false}

Log file location

<workdir>/.../LOG_DIR/<harness>/sidecar-logs/api-calls.jsonl

Collected by oss-crs artifacts via the log_dir field, and by CRSBench adapter to trial_dir/output/logs/crs/<crs>/log_dir/sidecar-logs/api-calls.jsonl.

Smoke test results

Tested against sanity-mock-c-delta-01 (fuzz_parse_buffer_section harness):

CRS	Bug-finding	Bug-fixing (with POV)
crs-bug-finding-claude-code	7 run-pov, all crashes	—
crs-bug-finding-codex	6 run-pov, all crashes	—
crs-bug-finding-copilot-cli	5 run-pov, all crashes	—
crs-bug-finding-gemini-cli	10 run-pov, all crashes	—
crs-claude-code	—	4 calls: build→run-pov(crash)→run-pov(no crash)→run-test(pass)
crs-codex	—	4 calls: run-pov(crash)→build→run-pov(no crash)→run-test(pass)
crs-copilot-cli	—	3 calls: build→run-pov(no crash)→run-test(pass)
crs-gemini-cli	—	3 calls: build→run-pov(no crash)→run-test(pass)

Test plan

• 22 unit tests for log entry construction and JSONL writing
• Smoke tested crs-bug-finding-claude-code, crs-bug-finding-codex, crs-bug-finding-copilot-cli, crs-bug-finding-gemini-cli in bug-finding mode (real Docker runs)
• Smoke tested crs-claude-code, crs-codex, crs-copilot-cli, crs-gemini-cli in bug-fixing mode with POV input (real Docker runs)
• Verified api-calls.jsonl persists in LOG_DIR on host
• Verified oss-crs artifacts reports log_dir containing the file

[azchin]

I'll reimplement this in the new builder sidecar workflow.