e2e harness proposals: shared classic-spec helpers, fewer Firefox restarts, clearer failure output, a truncated retried.jsonl, uncovered risky paths #266

Severity: low

• Duplication: inChrome, clickToolbar, suspendBackground, tabMode/modeOfTab, expectMode, hasOverlayPlayer and the "open the extension page, find the handle, set options, find the site handle" prologue are copied into eight classic specs; openExtensionPage exists four times; serveFile/range parsing is reimplemented in source-length, mp4-files, mp4-loading and mp4-fragmented-end (source-length's throws synchronously on bytes=5-2 via createReadStream ERR_OUT_OF_RANGE, which takes the worker down). A tests/e2e/classic-helpers.mjs and reuse of serveFile.mjs would make the next "found by URL, not last handle" fix land once.
• Speed: the extension suites start a fresh Firefox and reinstall the add-on per spec file (18 + 10 + 1 files, twice for the two builds). Grouping the short, state-free ext specs (background-log, capabilities, extension, download-names, download-blob-lifetime, header-rules, keybinds-storage, options-mpv) into one file, or raising maxInstances to 2 once the ports are unique, would cut the ext job noticeably; the web suite's per-spec servers could all live in onPrepare on one port with path routing.
• Clearer failures: playback.e2e.mjs and mp4-files.e2e.mjs log state, but expectEveryHlsSegmentListed's wait swallows its timeout (.catch(() => {})) and the following toEqual prints two arrays with no "audio playlist never arrived" reason; hls-live.e2e.mjs and mp4-loading.e2e.mjs have waitUntils without timeoutMsg; subtitle-search.e2e.mjs has waits with no timeout or message.
• logs/retried.jsonl is never truncated at the start of a local run, so a developer's summary accumulates every retry since the last rm -rf logs; clear it in each config's onPrepare or key it by run id.
• Uncovered risky paths named in CLAUDE.md: a saved profile in the real extension's chrome.storage for the keybind migration, the native host surviving the kill-on-close job object (hand-checked), and the OPFS 30 s "crashed worker" fall-through (storage.e2e.mjs stubs setup() but never a worker that stops answering).

Confidence: high.

From a line-by-line code review of the tree at 41363fd (2026-10-02). Each claim above was checked against the source before filing; the reproduction steps and line references are as of that commit.

Generated by Claude Code