[thekevinbot]

Fixes #1198

Five speed-of-light e2e tests, one per article use case, plus the embeddings one in the plugin suite. Each test builds its corpus in a temp dir, grows it by doubling until the fastest native pipeline takes at least a second, runs native and dirsql three times each, and asserts two things: the rows are identical, and dirsql wall time minus measured launcher startup (ten --help runs averaged) is within 1.1x native.

All five rows checks pass. All five timing checks fail today, each for the reason the test asserts, and are marked expected failures. Measured on this branch (32 cores, best of three; dirsql minus startup, native best of three):

test	corpus at the 1 s bar	native	dirsql	ratio	why
dossier (use_case_dossier_test.py)	400k markdown files	0.66 s (3 x find)	11.2 s	17x	three CTEs over ./ walk the tree three times, and one walk is ~4x find at this size (#1201)
notebook (use_case_notebook_test.py)	160k sections, one file	1.18 s (parser + sort/head)	1.85 s	1.6x	ingesting the parser's JSON rows is slower than json.load + sort (#1202)
join (use_case_join_test.py)	20k paper folders	~1 s (find x2, awk, sort, join)	>60 s (timed out)	-	unindexed nested-loop join, #1196
transcripts (use_case_transcripts_test.py)	1,024 JSONL files, 100 MB	1.41 s (one parser process)	52.0 s	37x	one parser process per file (#1203)
embeddings (plugin speed_of_light_test.py)	32k abstracts	1.31 s (model2vec batch encode)	4.46 s	3.4x	one worker round trip per value (#1204)

The grow until native >= 1 s rule sizes every corpus on the machine running it; there is no fixed millisecond number anywhere. A dirsql run more than ten times over budget ends its best-of-three early, since noise cannot account for an order of magnitude, which keeps the suite's red runtime bounded (about six minutes for the five tests here).

Each test carries pytest.mark.xfail(strict=True) with its gap as the reason, so the suite is green and attestation writes a receipt on any branch. Strict mode fails the suite the moment a test clears the bar with the marker still present, so each fix PR removes its marker in the same change. Epic #1205 tracks the five fixes, one PR each.

E2E Verification

• Ran e2e suites locally for every affected SDK
• Python SDK e2e: pass (38 passed, 4 xfailed in 3 m 56 s); plugin e2e: pass (17 passed, 1 xfailed in 24 s)
• TypeScript SDK e2e: N/A
• Rust core e2e (if applicable): N/A
• packages/python/e2e-attestations/<branch>.json receipt written if packages/python changed: not required, test-only diff
• packages/ts/e2e-attestations/<branch>.json receipt written if packages/ts changed: N/A
• Command(s) run: cd packages/python && uv run python -m pytest tests/e2e/ -q; cd plugins/dirsql-plugin-embeddings && uv run --with-editable ../../packages/python python -m pytest tests/e2e/ -q
• Result summary: five new tests xfail at the speed assertion with rows identical; existing suites green

Changelog / Migrations

• Changelog fragment: not required, test-only diff (the gate exempts tests/)
• Migration fragment: not required