feat(studio): gzip run-log commits, behavior_v2 default, bridge log-level ack (PR 2 of 3) — Studio v0.72

[mbreiser]

Studio side (PR 2 of 3) of docs/development/runlog-behavior-v2-plan.md. Stacks on #183 (the bridge PR); the base is that branch and should be retargeted to main once #183 merges.

Merge order note. Once this is live, new course runs commit as .jsonl.gz, which the analysis dashboard and the replay viewer cannot open until PR 3 (readers) lands. Recommend merging PR 2 together with or after PR 3. Until then a .jsonl.gz opens with gunzip or pixi run bridge -- --convert.

What changed

arena_studio.html v0.72

• Run logs commit gzipped. commitRunLog gzips the bridge export in the browser (GH.gzipBytes, native CompressionStream) and commits runlogs/<bench>/<name>.jsonl.gz through GH.commitFile, which uses the Contents API up to 30 MiB and the Git Database API above it (the Contents API rejects ~35 MiB+ files, which is what killed the 51 MB rig03-sr run). The log line, the upload modal and the run-summary line show raw → gzipped size and, when used, the large-file path. A browser without CompressionStream falls back to raw .jsonl.
• File ▾ → Run logging offers behavior_v2 (compact — default), behavior_v1 (legacy) and full. Nothing stored → v2; an explicitly stored v1 is honored (one release, per the plan).
• The bridge confirms the log level. At run start the Studio asserts the level, waits ≤ 800 ms for the bridge's log_control_ack, and names the confirmed level in the banner and the recorded transcript. When the bridge cannot write it: "⚠ bridge too old for behavior_v2 — logging behavior_v1 (restart pixi run bridge …)"; when it never answers (pre-3.0 bridge): "did not confirm … assuming behavior_v1". The Console's read-only #cFtLogLevel mirror follows the ack with a ⚠ and an explanatory tooltip on mismatch.
• run_metadata gains log_format = the acked level, or the inferred one for a non-acking bridge (behavior_v1 when v2 was requested, else the request itself). Set from the ack, never injected by the bridge, so the v1↔v2 file round trip from feat(bridge): behavior_v2 run-log format, --convert CLI, log_control ack (PR 1 of 3) #183 stays exact.

js/fictrac-bridge-client.js

• LOG_LEVELS = ['behavior_v2','behavior_v1','full'], default v2 (setLogLevel whitelist).
• Handles hello_ack → bridgeInfo, bridgeSupportsLevel(level) (null when unknown = old bridge) and log_control_ack → ackedLogLevel (cleared by every setLogging() and on close).
• New 'loglevel' event {source, requested, level, ok, levels, enabled, file} plus an 'err' log line on mismatch; waitForLogLevelAck(ms) resolves the acked level or null.

js/studio-github.js

• gzipBytes(input), isGzip(bytes).
• Git Database request builders reqCreateBlob / reqGetCommit / reqCreateTree (path allowlisted) / reqCreateCommit / reqUpdateRef (fast-forward only), directCommitLarge (repo → ref → commit → blob → tree → commit → ref, per-step error reporting), commitFile (size router; LARGE_FILE_BYTES = 30 MiB; thresholdBytes test hook; result carries via + bytes).

Docs: release notes v0.72, cshl-pipeline-test-plan.md size-limits section, CLAUDE.md Studio rule (readers must inflate; log_format only from the ack), plan doc status + "PR 2 implementation notes".

Tests

• tests/test-fictrac-bridge-client.js: 42 → 71 checks — v2 default, v1/full selectable, unknown ignored; hello_ack recorded + bridgeSupportsLevel; log_control_ack sets ackedLogLevel, resolves waiters, emits 'loglevel'; disabled ack; mismatch → ok:false + err line naming both levels; hello from a bridge lacking the level warns up front; close resets everything and resolves pending waiters null; old-bridge timeout path.
• tests/test-studio-github.js: 118 → 160 checks — gzip round trip against zlib.gunzipSync (UTF-8 preserved), gzip magic; builders' URLs/bodies/token containment, tree refuses a non-allowlisted path; directCommitLarge exact 7-call sequence with a mocked fetch (tree on the head root tree, commit parents = head, ref moved without force), blob-failure step reporting; commitFile small → Contents path, over threshold → git-db, UTF-8 byte counting; and assertions on the Studio HTML that commitRunLog gzips, names .jsonl.gz, routes through commitFile, keeps the raw fallback, offers the v2 option and stamps log_format.
• Full pixi run test passes; the four touched JS files are Prettier-clean; all five inline Studio script blocks parse.
• Not tested in a browser or on a rig (no hardware for a few days). The browser-only pieces are CompressionStream (Chrome ≥ 80; same API works in Node 26, where the gzip test runs) and the Studio DOM wiring, which is asserted textually. Bench check: recorded 20 s run → .jsonl.gz committed with the size shown and the banner naming behavior_v2; then run against an old bridge checkout and confirm the "too old" warning + log_format: behavior_v1 in the file; force the Git Database path by temporarily lowering LARGE_FILE_BYTES and confirm the blob/tree/commit/ref sequence lands one file.

🤖 Generated with Claude Code