Land Epoch Reviewer stack, remove stale docs/dead code, gate the release pipeline

[jdpigeon]

Four commits. The first recovers work that was stranded; the rest clean up after it, close the gap that made it possible to miss, and act on review.

Net: +2,320 / −1,011 across 42 files.

1. Land the Epoch Reviewer stack — 387fac2

PRs #222–#225 all report MERGED, but each merged into its parent branch, not main. Only the bottom of the stack (pyodide-analysis-fix, #220) was retargeted; when it landed, the children were never re-pointed. The result: main got a plan document and nothing else, while the implementation sat on epoch-reviewer-phase1, 24 commits ahead and invisible.

Clean merge, no conflicts — the stack never touched package.json/package-lock.json, so main's dependabot bumps didn't collide.

• Phase 0 (Epoch reviewer Phase 0: ship epoch arrays to a static Canvas renderer #223) — epoch arrays shipped to a Canvas renderer via transferable buffers
• Phase 1 (Epoch reviewer Phase 1: interactive rejection + live ERP + apply/save #224) — interactive rejection, live ERP, apply/save
• Phase 2 (Epoch reviewer Phase 2: bad-channel flagging, auto-flag suggestions, condition labels #225) — bad-channel flagging, auto-flag suggestions, condition labels
• Fix broken Clean→Analyze round-trip: persist cleaned epochs to host disk #222 — Clean→Analyze round-trip: persist cleaned epochs to host disk

Plus a fix cluster that matters independently of the feature:

Commit	Fix
beb1620	Scale Muse µV→V on load, so auto-flag stops rejecting every epoch
7fdccfb	Default EEG on + warn before a run with no brain data
c1cb40d	Remove the Emotiv-era USB receiver step from the Muse connect flow
69396e4	Show real condition names (Faces/Houses) instead of STIMULUS_n
68a410b	Open workspace folder via absolute path in the main process

2. Remove stale docs and dead code — a97bd04

Dead code, all verified zero-caller after the merge:

• plotTestPlot — debug scaffolding that rendered a four-point zigzag titled "Test Plot" into the real plotKey: 'topo' UI slot. Its only caller was swapped for plotTopoMap in Fix Pyodide analysis pipeline (harvest #194: dataKey routing, MEMFS, MNE 1.x) #220.
• cleanEpochsPlot — rendered a matplotlib epochs figure in WASM then immediately plt.close()d it, never routed to the UI. Became dead in this merge — the Epoch Reviewer replaced its last caller, as the plan predicted.
• parseSingleQuoteJSON — obsoleted by the worker's toJs() dict_converter.

Docs: deleted docs/lsl-implementation-plan.md (477 L — LSL shipped in #204/#221; every file in its "New files to create" inventory exists). Corrected device-connectivity.md, which opened with "currently: Muse only" — false since the EEGDriver registry.

Also dropped the "Pre-existing TypeScript errors (do not treat as regressions)" entry from .llms/learnings.md and TODOS.md. Those errors are fixed, and a standing instruction to ignore type errors in two files is a permanently disabled alarm.

3. Gate the release pipeline — f3f432c

The Release workflow has never executed. No tag has ever been pushed. As written, a tag ran npm install + package-ci and published whatever compiled — no test, lint, or typecheck on the release path.

• New verify job (test + typecheck + lint) that the publish matrix needs:. Nothing builds or uploads until it's green.
• Split package-ci into build → verify artifacts → publish, so a build with a missing Pyodide payload can't be published.
• npm ci in both workflows, so the tree CI passed on is the tree that ships.
• fail-fast: true to limit half-good uploads across the three OS legs.

⚠️ Behavior change: publish.releaseType: "draft". electron-updater is wired to this repo, so a tag would otherwise push an update to installed clients the moment it built. Draft means artifacts upload but a human publishes.

4. Address review — 2b339c7

−1,490 net. Every item from the review was taken; two were taken further than proposed.

Docs + the vendored skill (−950). Deleted faces-houses-qa-plan.md (a manual QA script nothing reads), epoch-review-ui-plan.md (a plan for work in this diff; Phase 3/4 already live in TODOS.md), and the accidentally-vendored agent-browser skill including its checked-in symlink.

The path contract: detect → make impossible. The review was right that regex-scraping src/main/index.ts to compare a folder name is brittle. New src/shared/pyodideAssets.ts exports PYODIDE_SOURCE_DIR / PYODIDE_RESOURCE_DIR; src/main/index.ts imports them, and tests/pyodideAssets.test.ts is deleted. One assertion survives in build.check.ts: package.json is JSON and cannot import, so extraResources ↔ constants is the one pairing a shared symbol can't enforce.

Test harness collapsed. build.test.ts → build.check.ts — the default glob never matches it, so vitest.build.config.ts and the vitest.config exclude both go. Two files coordinating to hide one test, replaced by a filename.

YAGNI. getPtpThresholdPreset(_device) ignored its parameter and returned a literal → PTP_THRESHOLD. hideAutoFlagEpochs (no experiment sets it) and the always-true showAutoFlag param. EpochArraysMeta.sfreq/.drop_log and SuggestedRejection.channel/.peak_uv — zero readers outside fixtures, dropped from utils.py too. The amp ± gain control (lanes already autoscale). The settling banner (HelpSidebar says the same thing, and this PR adds the Help entry point to that screen).

Native + stdlib. role="button"/tabIndex/onKeyDown → <button type="button"> in both places. times.findIndex((t) => t >= 0). conditionIndexForCode → Math.max(0, codes.indexOf(code)) inline. The channel-count warning now uses showMessageBox, matching handleCleanData, instead of a Radix Dialog plus a state field.

Shrink. downsampleMinMax's width >= length branch, meanTrace's empty guard, the double-modulo in conditionPalette, utils.py's single-file short-circuit, and the dataKey === 'epochArrays' special-case in the worker. palette is now required in plot_topo/plot_conditions, so the copied _DEFAULT_CONDITION_PALETTE is gone.

One deviation worth flagging. The review asked to drop peak_uv from utils.py, but test_ptp_in_physiological_range — the µV→V regression guard the review explicitly kept — read its value. Rather than keep the field or lose the guard, the test now computes peak-to-peak from the epochs directly. It tests the data instead of the reporting payload, which is what it was always trying to do.

Test plan

Authored on a machine with no Node, no node_modules, and no MNE — CI is the first execution. The previous push already answered the biggest unknowns: the ubuntu leg passed, which means npm ci succeeded (lockfile in sync), the merged stack passes against main's dependabot bumps, and test:build ran clean after Build.

This push needs re-verification of:

• npm test — fixture edits in EpochReviewer.test.tsx / epochArrays.test.ts, and the removed conditionIndexForCode / downsampleMinMax cases
• npm run typecheck — the PTP_THRESHOLD swap, dropped interface fields, and the renderReview signature change
• npm run lint — two <div role="button"> → <button> conversions
• npm run build + npm run test:build — the renamed check and new shared constants
• Analysis (ERP round-trip) — test_suggest_rejections.py rewritten; palette now required in plot_topo/plot_conditions

Not covered by CI:

• Manual: run Faces/Houses, Clean, reject epochs, confirm the live ERP updates and the cleaned .fif round-trips to Analyze
• v1.0.0-rc.1 dry-run — the release path is still unproven end to end

Still open

Recorded in TODOS.md, not addressed here:

• No packaged-app launch smoke test. build.check.ts proves the payload is on disk; nothing proves the app starts. That's the gap that would catch a prod-only Pyodide regression automatically.
• The release pipeline is gated but unexercised. Worth an RC tag before any demo depends on a released build.