[OPIK-8021] [QA] Proposed e2e specs for the virtualized traces table (from #7947 exploration) - #7970

Generated by the release QA side flow — draft, needs human review before merge.
Nothing here has been reviewed by a person. Two specs were written from an
exploratory testing pass and run green against the source PR's own deployed
environment; a third candidate was dropped (see below).

Where these came from

Exploratory testing of #7947 — "OPIK-8021 [FE] perf: virtualize
the traces table's columns and rows" — against that PR's own deployed
environment (https://pr-7947.dev.comet.com, serving 2.2.37-7947-merge-3045,
OSS install, workspace default). A human worked the flows by hand there first;
these specs automate the two that held up.

This PR targets andriid/OPIK-8021-traces-table-columns-perf, not main.
The behaviour asserted here — a traces table that renders a moving window of
rows and columns instead of all of them — only exists on that branch. On main
the table still renders every row, so the specs' own precondition checks
(fewer rows in the DOM than were seeded, spacer cells present in the header
row) would fail. Rebase this onto main once #7947 lands.

Written and verified against 4b66cdd4c1d0583a08ffd574fb983371c40bba8b, which
was #7947's head at the time.

The specs

Both live in tests_end_to_end/e2e/tests/trace-explore/traces-table-virtualization.spec.ts.

1. @cap:traces.list-traces — row window renders every trace exactly once

Seeds 60 traces, reads the API's ordering for the same project, then sweeps the
page body from top to bottom in 250px steps. At every offset it asserts the
rendered data-row-ids are a contiguous slice of the API's ordering starting
where they claim to, and that no id is in the DOM twice at once. Across the whole
sweep, every seeded trace must have been seen, and the final window must be the
tail of the ordering.

The failure this is for is a row silently dropped or repeated at a window seam —
which renders as a plausible-looking table, not as an error. traces.list-traces
is nominally covered by trace-explore-smoke.spec.ts, but that spec seeds three
traces and so never crosses the 25-row threshold; this is the only spec that
reaches the windowed code path at all.

Verification: passed. npx playwright test tests/trace-explore/traces-table-virtualization.spec.ts
against https://pr-7947.dev.comet.com, green on 4 consecutive runs (~24s each).

2. @cap:traces.configure-columns — column window keeps everything aligned

Gives the same 60 traces 45 distinct feedback-score names, which the Logs table
offers as 45 extra columns (67 in total), then turns every column on from the
Columns picker. At five horizontal scroll offsets it reads the header's column
sequence — including the aria-hidden spacer cells, which are positions in the
row like any other — and asserts every rendered row's cell sequence and the
<colgroup> match it position for position, and that the sampled sequences
are not all identical (i.e. the window actually moved). Finally, at maximum
scrollLeft, the pinned select column must be position: sticky at left: 0
and geometrically flush with the scroll container's left edge, on the header and
on every rendered row.

The failure this is for is a spacer of the wrong width shifting every cell one
column across: correct data under the wrong header, which looks like correct
data. traces.configure-columns had no spec at all before this.

Verification: passed. Same command and environment, green on 4 consecutive
runs (~16s each).

What this change also touches

• fixtures/windowed-traces.fixture.ts (new) — windowedTraces seeds 60
  traces in one batch write with caller-minted UUIDv7 ids one second apart, then
  polls until the project lists exactly those ids and returns them in the API's
  order. wideTracesTable chains from it and adds the 45 score names. Sixty
  rather than the forty the exploration used, so "fewer rows in the DOM than were
  seeded" has ~20 rows of headroom rather than 2 and does not depend on the
  runner's viewport height. Teardown deletes the traces explicitly — deleting a
  project does not cascade to them, and global-teardown's prefix sweep only
  knows about experiments, datasets and projects.
• pom/logs.page.ts — window-reading helpers (header/cell/colgroup
  sequences, scroll extent, a scroll-and-settle step, pinned-column geometry) and
  the Columns picker. No raw selectors leak into the spec.
• core/backend/client.ts — createTracesBatch and scoreTracesBatch. The
  Python bridge creates one trace per call and flushes each time, which is right
  for a three-trace fixture and not for sixty traces carrying 2,700 scores.
• coverage/taxonomy.yaml — spec added to the traces area's specs: list;
  configure-columns flipped to covered: true, tier: t2-cuj. list-traces was
  already covered: true at t1-smoke and is left as-is.

Two things a reviewer should look at

1. The scroll container has no data-testid. PageBodyScrollContainer.tsx
   renders the element the table virtualizes against, and its only distinguishing
   marks are Tailwind classes. The POM resolves it as the nearest scrollable
   ancestor of [data-table-wrapper] — semantic rather than class-coupled, and
   the same lookup the virtualizer makes through React context. The house
   conventions say to add a data-testid in the same change; that is deliberately
   not done here, because the target environment serves a prebuilt frontend
   image, so a testid added in this PR could not have been exercised by the run
   above. A data-testid="page-body-scroll-container" would be a strict
   improvement and the POM helper should switch to it.
2. The exploration flagged a possible gap in the gate, which these specs do not
   assert. TracesSpansTab.tsx builds one virtualization object for both
   axes and the comment says one decision drives both, but the column axis was
   observed not to engage on a wide table with few columns. That is an
   optimisation gap, not a correctness bug — nothing renders wrongly — so no
   assertion here pins the current behaviour. Worth an author sanity check
   independently of this PR.

What was deliberately not written

One candidate of three was dropped: select-all and bulk delete covering rows
outside the rendered window (traces.delete-traces). It verified by hand, but
traces.delete-traces is already covered by trace-delete.spec.ts, and the
right shape is probably to raise that spec's seed count past the windowing
threshold rather than add a second delete spec — a call for whoever owns that
file, not for a generated PR.

Verification, in full

Run from tests_end_to_end/e2e/, against OPIK_BASE_URL=https://pr-7947.dev.comet.com
(OPIK_DEPLOYMENT=oss, workspace default):

npx tsc --noEmit                                                    # clean
python3 tests_end_to_end/coverage/tag_lint.py \
  --taxonomy tests_end_to_end/coverage/taxonomy.yaml \
  --estate tests_end_to_end                                         # 44 specs, 0 problems
npx playwright test tests/trace-explore/traces-table-virtualization.spec.ts
                                                                    # 2 passed
npx playwright test tests/trace-explore/traces-table-virtualization.spec.ts --repeat-each=3
                                                                    # 6 passed
npx playwright test tests/trace-explore/                            # 20 passed

The whole trace-explore/ directory was run because this change touches a shared
POM, a shared fixture chain and the shared backend client.