Add behavioral projection helpers to testUtils (projectColumn,
projectRenderResult, rowDigests, hunkDigest, patchDigest,
annotationProjection, collectRowSourceMismatches, countDeclaredRows)
so tests can assert just the rows/indices/types/text they own instead
of snapshotting entire render results. Remove findBufferElements: its
data-virtualizer-buffer attribute only ever exists on live DOM nodes,
never in renderer HAST, so every assertion built on it was vacuous.

Rewrite DiffHunksRendererVirtualization.test.ts:
- 1.1/1.2 assert the renderer's bufferBefore/bufferAfter passthrough
  instead of the impossible buffer elements, and document the row-count
  derivation (514 hunk rows + 3 auto-expanded single-line gaps = 517).
- 3.5 now proves the a9ff17b7 regression guard with a differential
  oracle (windowed render === rows 30..79 of the full render) plus an
  explicit containment check on the expanded indices 57-76, instead of
  a 2.8k-line snapshot.
- 6.1/6.2 finally do what their names promise: rendered window rows are
  compared to the full render slice and their text to the exact source
  lines on the correct side.
- Delete test 2.1 (the 517 count was pinned three times) and 7.2 (after
  removing its redundant full render it duplicated 7.1 verbatim);
  simplify 4.4's dead else-branch and 7.1's decorative toBeDefined.
- Correct the stale fixture geometry comments (the header undercounted
  rendered rows and misnumbered hunk ranges) and document that
  startingLine/totalLines index renderable rows, not virtual indices.

Deletes the 2.1MB snapshot file; the remaining assertions are exact.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

DiffHunksRender.test.ts: the buffer-prepending tests now assert the
named behavior directly — buffer rows appear in the correct column
with sizes equal to each change block's add/delete surplus, derived
from the parsed hunkContent — instead of burying it in 2,300-line
full-HAST snapshots. All seven whole-result snapshots become compact
rowDigests projections (or are dropped where countInlineDiffSpans
already pins the behavior), and the five 'parsed diff' snapshots
become verifyHunkLineValues invariant checks.

annotations.test.ts: the three full-AST snapshots (151k lines frozen
per-token theme styles) become annotationProjection snapshots — the
ordered (line, annotation, slots) placement record the tests are
actually about; the precise loop assertions above them are unchanged.

patchFileRender.test.ts: the known-tricky-patch regression pin is now
a patchDigest + rendered-row projection instead of two whole-object
snapshots, with hunk invariants checked explicitly.

Also fixes an edge in verifyHunkLineValues surfaced by the new
invariant checks: a diff whose addition side is empty (file deleted to
nothing) has additionStart/additionCount 0, and the unclamped
lastHunkEnd of -1 invented a phantom trailing context line.

Snapshot churn: 264k lines of full-HAST output replaced by 265 lines
of reviewable projections.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

parseMergeConflictDiffFromFile: the three whole-result snapshots of
the 20,000-line conflict fixture (12.5MB, 607k lines — unreviewable
and update-blessed by definition) become explicit assertions at each
maxContextLines width: hunk invariants, diff sides exactly equal to
the conflict-free current/incoming texts, no markers on either side,
44 actions (one per fixture conflict), monotonic hunk-row growth with
wider context, plus a compact per-width geometry digest. The test name
no longer promises "timing" it never measured (the benchmark script
owns timing).

parsePatchFiles: the 94k-line whole-result snapshot of the 400KB
diff.patch becomes a one-line-per-hunk patchDigest snapshot; the small
targeted snapshots (final blank line, malformed patch) stay. Invariant
checks now assert .errors toEqual([]) so failures print the actual
errors instead of valid === false.

parseDiffFromFile: full-result snapshot becomes a hunkDigest snapshot.

hunkDigest now renders each hunk as a single line, keeping the 400KB
patch's digest reviewable (~1.5k lines instead of 7.4k).

test/__snapshots__ total: 966k lines / 20MB before this series, now
4.5k lines / 156K.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

FileRenderer.test.ts keeps its full highlighted-AST snapshot and is
now documented as the suite's single full-fidelity token/style canary:
theme or tokenizer changes should churn exactly this one snapshot,
which at 1.6k lines is small enough to review line by line.

FileRenderer.ast.test.ts: the css check accepted any string while the
non-worker path hardcodes '' — assert that contract explicitly; the
preNode test asserted no actual property — it now checks data-file,
data-overflow, tabIndex, and the gutter-width style derivation; the
totalLines test no longer re-asserts file2 (covered by the structure
test above it).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Cannot-fail assertions, each replaced with one that discriminates:
- CodeView.collapsed: top <= bottom on window specs is guaranteed by
  createWindowFromScrollPosition's clamping; now asserts the clamped
  window reaches the list tail and scrollTop stays within the derived
  collapsed content height. Also drops the sticky-specs test whose
  expected value was read from the instance under test.
- CodeView.pointerEvents: the "cancels pending restore work" half
  could never fail (post-cleanUp restores are no-ops); renamed to what
  it verifies and dropped the dead post-wait re-assertions plus the
  incidental intermediate-state pins.
- hydration: not.toBe('0px') passed with no placeholder at all; now
  requires the placeholder and asserts its exact height derived from
  DEFAULT_VIRTUAL_FILE_METRICS and parsed row counts.
- InjectedRowHooks: isHastElement on a helper that only returns hast
  elements replaced with exact action-button table (action ids,
  conflict indices); rowCount now derived from parser output instead
  of the rendered result itself.
- CodeView.rangeScroll: line/range equivalence now anchors on the
  exact derived center position so both paths silently no-oping at 0
  can no longer pass.

Magic numbers replaced with derivations:
- iterateOverDiff: 517/490 become countDeclaredRows(diff, style)
  (hunk-declared rows + auto-expanded gaps at the collapsed-context
  threshold); fixes the split test collecting unifiedLineIndex (a
  copy-paste bug nothing caught because it was never asserted) and
  adds a metadata-derived split-pairing assertion; deletes the shallow
  big-fixture expansion check superseded by the synthetic-fixture
  tests.
- computeEstimatedDiffHeights: every raw expected number (36/64/.../
  1454) rewritten as arithmetic over the named metrics and parsed
  fixture geometry, in the style of virtualFileMetricsPadding.
- CodeView.scrollAnchoring: verbatim 2,000,000/12,000,000px pins
  become named mirrors of the source rebase constants with
  relationship assertions (delta advancement, threshold-derived
  bounds).
- virtualizedFileDiffEstimatedHeights: inline 5_000 named as
  LAYOUT_CHECKPOINT_INTERVAL pointing at the source constant.

Truth-in-naming and accuracy:
- CodeView.updateItemId now asserts the rename's positive effects
  (same element/instance under the new id only, retargeted pending
  scroll actually scrolls).
- FileDiff.partialRender and sharedHighlighter tests renamed to what
  they verify; advancedStickySpecs window-helper comments un-inverted.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Document the shared harness/projection conventions, the single-canary
snapshot policy, and the coverage gaps confirmed by the test audit
(worker pipeline, SSR hydration round trip, ScrollSyncManager, real
Virtualizer, drag selection, shiki-stream, and friends) so the next
person touching those areas knows the tests are missing rather than
assuming they exist.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>