# NOTE2 verification

The final head is f10c311fb6743847d91d7195ecd03232e7aa0cf2. The baseline is origin/master d5022199ca5adecca8cf2fbc69fe013820ebd0fa. Tests run with Node v22.23.2 on Linux in the worktree's isolated Obsidian vault. No shared vault was used.

## Reproduction and root cause

The original code inserted captures through the note editor. In Live Preview, Obsidian's cellSanitizerPlugin intercepts this edit and asynchronously maps it into a nested cell editor. That editor contains decoded pipes, while the incoming change assumes the Markdown representation's length. After the first escaped pipe, a later capture raises `RangeError: Applying change set to a document with the wrong length` and does not persist.

`red-e2e.log` records the regression failing on master: the disk contains only `anchor00:00:30 \| cut`, while the assertion expects the following 10-second segment too. `before-errors.txt` records the RangeError. The initial diagnostic run in `red-e2e-stale-diagnostics.log` is excluded because it detected errors left by a manual probe.

The fix resolves `editor.editorComponent.tableCell.editor` afresh for each capture, validates its writer, and keeps its receiver. It removes the Markdown escape before each pipe for the nested editor. Obsidian encodes the cell when writing it back to the note. Snapshot Undo still compares the whole note. For an active cell, Undo extends the inverse start to the line boundary so Obsidian does not intercept and remap it. The end stays unchanged so a follow-up capture remains in the cell.

The real player-button test also exposed a focus race. Obsidian clears activeEditor before its deferred active-leaf-change event reaches the service. The trace in player-call-probe-2.log shows the correct active leaf still held by the service while a capture went to the episode note. The fix consults that validated leaf before the saved last editor. red-deferred-focus-unit.log reproduces the wrong destination; the fixed test passes.

## Runtime evidence

The ten focused tests in `tests/e2e/table-cell-capture.test.ts` pass on both Obsidian 1.14.4 and 1.13.7:

- Timestamp, 10-second segment, 20-second segment, exact saved-note Undo, then another capture.
- Full player, mini player, and the installed MediaSession previous-track callback.
- Honest Undo refusal after another cell changes.
- Exact Undo when the column width stays constant.
- Existing pipe-escape parity.
- A paragraph after leaving the table, preserving literal pipes and newlines.
- Undo with another view of the same note, including an active cell in the second pane.
- Undo after closing the captured tab.
- Source mode captures and exact Undo.

The headset test records and invokes the real registered callback. It does not simulate physical headset hardware.

Automated journeys use the local `Table capture podcast` feed and `Table episode`, paused at 30 seconds, with `{{time}} | cut` followed by a newline. The fixture server chooses its port and the repository resolver maps podnotes.test to localhost. The before/after screenshots use `http://podnotes.test:33494/feed.xml`, `Fixture Episode 1`, and `{{linktime}} | cut` followed by a newline. `after-note.md` contains all three link URIs, with escaped pipes. `after-errors.txt` says `No errors captured.`

Before: https://files.bagerbach.com/before-table-g1jj44vpklv6.png

After: https://files.bagerbach.com/after-table-gjdltph1vimq.png

## Mutations

Run `python3 .obsidian-e2e-artifacts/note2/mutate-unit.py` and `python3 .obsidian-e2e-artifacts/note2/mutate-runtime.py` from the worktree root. Both restore source in `finally`. Rebuild the correct bundle afterwards.

| Wrong implementation | Decisive observed failure |
| --- | --- |
| Use the outer editor for captures | Runtime repeated-capture test saves only the first capture and records the RangeError. Unit cell text remains `anchor`. |
| Pass escaped Markdown directly to the cell | Runtime test sees literal `\|` instead of `|` in the cell editor. Unit parity test also fails. |
| Use a narrow inverse for Undo | Runtime wide-column test leaves `anchor00:00:30 \| cut` in the saved note instead of restoring `anchor`. It also records the RangeError. Unit repeated-capture Undo leaves the second segment. |
| Extend the inverse end to the row boundary | The repeated-capture runtime test restores the note but the next capture lands outside the cell. |
| Ignore the pending active leaf | The unit test returns destination `note` and the episode-note path instead of destination `editor` and `Journal.md`. |
| Drop the cell writer's receiver | Unit capture rejects with `Cannot read properties of undefined (reading 'text')`. |
| Omit the null guard | Paragraph capture leaves `Notes` unchanged and rejects with `Reflect.get called on non-object`. |
| Remove all backslashes before a pipe | Unit parity test loses the existing pair of backslashes. |

Every runtime mutant failed its intended content assertion. A nonzero exit code alone was not counted as evidence.

## Checks and size

`npm run lint`, `npm run format:check`, `npm run typecheck`, `npm run build`, `npm run test`, `npm run test:tooling`, `npm run docs:build`, and `npm run size` pass. There are 2,547 application tests and 13 tooling tests. The docs build checks 18 pages and all 112 legacy fragments.

The full production bundle changes from 311,100 to 311,433 raw bytes, an increase of 333. Gzip at level 9 changes from 98,345 to 98,473 bytes, an increase of 128. The gzip gate is 100,000 bytes, leaving 1,527 bytes.

An independent verifier ran the full suite against final head `f10c311fb6743847d91d7195ecd03232e7aa0cf2` on a fresh isolated vault for each runtime. Both runs used Node v22.23.2 and `TMPDIR=/data/home/tmp/podnotes-note2`.

| Runtime | Command | Result |
| --- | --- | --- |
| Obsidian 1.14.4, installer 1.14.4 | `fnm exec --using 22 -- npm run test:e2e` | 24 files, 186 tests, all passed |
| Obsidian 1.13.7, installer 1.13.7 | `PODNOTES_OBSIDIAN_APP=/opt/Obsidian/obsidian fnm exec --using 22 -- npm run test:e2e` | 24 files, 186 tests, all passed |

Each run includes all ten table-capture cases. No tests failed or were skipped. The runtime error recorder reported no uncaught errors. The final `main.js` SHA-256 is `6e4fd5e80868e5254c37582bb4a8bd690a1b1b36fe52a5889bf47b12041df8bc`.

The verifier stopped each instance. Both explicit cleanup dry runs report `pids: none`. The tree was clean before and after verification. The text logs, exit statuses, version records, and cleanup records are included in the evidence archive.

## Issue 390

Already fixed by #468. The settings input derives its placeholder from DEFAULT_SETTINGS. The issue was closed with exact master-code references: https://github.com/chhoumann/PodNotes/issues/390#issuecomment-6163151736. This change adds no settings code.


## Review and dispatch record

The single interrogate pass accepted the scoped test-selector correction and added second-pane and post-Undo checks. The latter caught the inverse-end expansion defect. No findings remain unresolved. `reviews.md`, `review-claude.md`, and `decisions.tsv` record the accepted and rejected findings.

| Task | Provider and model | Result |
| --- | --- | --- |
| Investigation | claudeAgent / claude-opus-5-5 | Confirmed the nested cell editor mechanism |
| Design candidate | claudeAgent / claude-fable-5-1 | Compared cell writer and Undo choices |
| Design candidate | codex_pool / gpt-6-astra | Compared cell writer and Undo choices |
| Design judge | claudeAgent / claude-opus-5-5 | Selected the direct cell writer |
| Implementation | claudeAgent / claude-opus-5-5 | Delivered the scoped fix and tests |
| Comment review | claudeAgent / claude-opus-5-5 | Accepted comment and helper cleanup |
| Interrogate | codex_pool / gpt-6-astra | Found the unscoped test selector, fixed |
| Interrogate | claudeAgent / claude-fable-5-1 | Added probes that exposed post-Undo behavior, fixed |
| Final independent runtime verification | claudeAgent / claude-opus-5-5 | Both full runtime suites passed on the final head |

All task results were consumed and the delegated tasks were explicitly stopped. No further review or monitoring is active in this thread.

## Decisions and limits

Fix Root Causes led to direct writes through the decoded cell editor after reproducing the outer editor's length mismatch. Prove It Works required saved-file assertions in real Obsidian, independent final runtime runs, and mutations that fail for the intended content mismatch.

The fix uses a private Obsidian cell-editor API. Both supported runtime lines pass, but future Obsidian changes need the same regression journey. Headset coverage invokes the installed MediaSession callback and does not test physical hardware.

There are no settings, storage, public API, or URI changes. Published docs remain unchanged. The PR records the exact timestamp-documentation update for release. No unresolved correctness findings or additional implementation follow-ups remain.