[raiseCatError]

Stacked on #414 (/edit) → #413 → #412 → #411. Merge in order.

What

Deterministic, post-completion readings of recognized output (src/output/semantic.ts). Each recognizer accepts only its own unmistakable evidence and fails closed; counts are the ones the tool printed, and items are verbatim lines (at most 20, plus a count).

Recognizer	Evidence
Node's test runner	spec ℹ tests/pass/fail or TAP # tests/pass/fail; failures from the failing-tests section or not ok (TODO/SKIP excluded)
Jest, Vitest, pytest	their own summary lines; ● / FAIL / FAILED items
cargo test	test result: lines, summed per binary
go test	only from go test; ok / FAIL package lines and --- FAIL: tests
TypeScript, Rust, GCC/Clang	their diagnostic formats; only from build commands or failed commands, so a successful rg error is never read as a build

Presentation (semanticRows.ts, derived rows like fold hints)

• Overview row: one row in the block's output area, e.g. ✔ node --test · 1,772 passed · 0 failed · 25 skipped. Safe glyphs +/x, and readable with NO_COLOR.
• Details row: shown when failures or diagnostics are listed. It starts open when the raw output is folded (that is what the fold hides) and closed when the output is expanded. Click the row, or use Actions → Show or hide details. The reader's choice is kept per block, independent of Ctrl+O.
• When nothing is shown: short passing output (under 10 lines), unrecognized output, Ask blocks and agent blocks.
• Never touched: stored transcript lines, raw folding, find, drag-selection and plain /copy. Derived rows are skipped by find and selection. Readings are cached per record and never persisted; a resumed transcript reads them again lazily.

Copy

• /copy --raw (default), --semantic and --both. These are mutually exclusive, not allowed with --report or ui, and combine with every selector.
• --semantic and --both copy directly; without a selection they take the latest output.
• A record without a reading is copied raw, once, and the note says so.
• --both puts the raw output and the reading under labelled ── NMSh · … ── separators.

Settings

Settings → Transcript → Semantic Output (On by default). It is also reachable from Setup's Transcript section.

Not in this phase

• External providers (jc, delta).
• A fully collapsed raw section.
• A separate semantic search scope.

These are recorded in the design doc.

Tests

• semanticOutput.test.ts (10 tests):
  • one realistic fixture per tool, plus fail-closed cases (incomplete summaries, look-alikes from non-build commands, unrelated output);
  • bounds and deduplication, and reading from the end of large output;
  • presentation: rows come before the output, stored lines are unchanged, search ignores them, details follow the raw fold until toggled, and Off draws raw only;
  • the usefulness gate, /copy flag parsing, and the semantic and both payloads.
• semanticOutputLive.test.ts: a real PTY draws the overview and the listing in the block; /copy --semantic copies the reading and /copy latest the raw output, through a stand-in clipboard.
• Updated audits: block-action registry, and Setup coverage for the new row. 154 output, settings and copy files plus 37 copy files are green.

Not verified

Physical terminal checks: docs/testing/v021-physical-qa.md, section 4, not run.

Refs #386, #410