[Goal] Execute scenario 7 against independent pandas snapshots

Start by inspecting tests/xval/runner.test.ts, especially assertStep and materializeSnapshotStep, then read golden/generate.py:scenario_7. Implement and integrate an independent executor for the ten scenario 7 steps using real tsb APIs, with strict comparisons and negative controls. Run bun test ./tests/xval/, typecheck, lint, the full tests, and pinned snapshot regeneration; done requires unchanged golden data and exact-SHA CI evidence.

[Goal] Execute scenario 7 against independent pandas snapshots

Goal

Turn the ten recorded steps of scenario_7 into real cross-validation of tsb operations. Today tests/xval/runner.test.ts constructs each "actual" result from that same expected snapshot and compares it back; this checks materialization, not the named operation. Replace that circular path for one complete, bounded scenario without attempting a broader pandas port.

This checkpoint establishes value, missing-value, shape, and ordered-label parity for scenario 7 only: 10 of the 60 recorded steps. It does not establish dtype parity, pandas-compatible in-place mutation, the other six scenarios, full API parity, or performance parity.

Completion Contract

1. Build original inputs from the literals in golden/generate.py:scenario_7, not from any expected result. Keep execution in a small function/module that has no snapshot argument, snapshot import, file read, expected-data lookup, or hardcoded expected result. Import the real public APIs from tsb:

   Snapshot steps	Execute
   1–2	combineFirstSeries(a, b), then combineFirstSeries(combined, c)
   3	dataFrameUpdate(df1, df2) and compare its returned frame
   4	whereSeries(s, value => typeof value === "number" && value > 20, { other: -1 })
   5	maskSeries(s, value => typeof value === "number" && value > 30, { other: 0 })
   6–7	Both results of alignDataFrame(left, right, { join: "outer" })
   8–9	Both results of alignDataFrame(left, right, { join: "inner" })
   10	combineFirstDataFrame(alignedLeft, alignedRight) using the actual outputs of steps 6–7

2. Compare those actual results against the unchanged, independently generated golden/snapshots/scenario_7.json. Assert exactly ten distinct step IDs, correct Series/DataFrame result kind, data values, missing positions, shape, column order, and index order with numeric versus string label types preserved. For example, step 3 retains numeric row labels 0,1,2,3,4; do not stringify both sides to make mismatches disappear. The existing 1e-10 numeric tolerance may remain; null/undefined/NaN may use the documented snapshot missing-value encoding. Do not normalize away nonmissing value or label differences.

3. Add at least two negative-control tests that run the same executor and comparison path but substitute deliberately incorrect operations: a no-op whereSeries and a no-op or wrong-join alignDataFrame. Each must cause the relevant snapshot assertion to reject the output. Test-only dependency injection with real APIs as the production defaults is acceptable; no global mock leakage or production-library modification. Returning expected snapshot data would bypass these faults and must make the negative-control tests fail. Include a comparator regression proving numeric label 1 does not equal string label "1".

4. Remove scenario 7's use of materializeSnapshotStep as its "actual" result. Preserve the other scenarios' existing fixture/materialization checks, but label those checks accurately and report the scope as one executed scenario / ten executed steps; fifty remaining steps are materialization-only. Do not relabel the other scenarios as verified operation parity.

5. Keep known semantic limits explicit in test names/comments and the completion report: several tsb outputs infer integer dtype where these pandas snapshots retain float dtype; dataFrameUpdate returns a new frame rather than mutating the source in place. Do not claim these differences are verified parity, alter the expected dtype metadata, force result dtype to match, or broaden this checkpoint to fix the library.

6. Required tests, strict typecheck, lint, and the existing CI must pass on the exact published candidate commit. Record that SHA, its CI run URL, positive and negative-control test names/counts, and the ten exercised steps. Snapshot regeneration with pinned pandas/numpy must leave the committed golden data unchanged. Pending publication, a green unrelated SHA, fixture self-comparison, or merely constructing Series/DataFrames is not completion.

Evidence / Verification

First reproduce the current circular path by inspecting assertStep and materializeSnapshotStep. Read golden/generate.py:scenario_7 and confirm the executor uses its original inputs and real intermediate results.

Run the focused positive and negative-control tests, then:

bun test ./tests/xval/
bun run typecheck
bun run lint
bun test ./tests/
python -m pip install pandas==2.2.3 numpy==2.1.3
python golden/generate.py
git diff --exit-code -- golden/generate.py golden/snapshots/

Use existing CI for pinned snapshot regeneration if Python is unavailable locally; keep completion pending until the exact-SHA run proves it. Attach a compact table of each scenario 7 step, invoked API, and asserted result. Report the negative-control failures as expected test outcomes, not as production test failures. Do not use or require a benchmark as evidence of correctness.

Scope and Constraints

Allowed changes are limited to tests/xval/**: a scenario executor, typed comparison helpers/fixtures if necessary, the replay integration, and positive/negative-control tests. Keep original input literals independently copied from the Python generator; expected outputs remain exclusively in the existing golden snapshot.

Do not edit src/**, golden/generate.py, golden/snapshots/**, README.md, .autoloop/programs/**, workflows, package scripts/dependencies, agent policy, branch rules, or credentials. No any, new as casts, @ts-ignore, runtime escape hatches, skipped tests, reduced tolerances, snapshot-result-derived inputs, or hand-authored fake tsb implementations. Preserve unrelated existing checks. This is test infrastructure for supported operations, not permission to port missing pandas features.

Iteration Policy

Choose one reviewable checkpoint: add the real scenario 7 executor and integrate its ten outputs, establish positive and negative-control evidence, and publish one canonical draft PR. If needed, separate the strict typed comparator and executor into small commits on the same branch. Record pending publication and verify the actual published SHA and CI on a later run before accepting completion. Report what was executed versus merely materialized; do not start porting scenarios 1–6 as part of this goal.

Blocked Stop Condition

Stop and report the exact failing step, observed versus expected result, command output, and smallest required maintainer decision if an existing tsb operation cannot satisfy this value/label contract without changing production code, if independent golden data appears incorrect, or if publication/verification cannot run. Known dtype and in-place-mutation differences are excluded explicitly above, not silently normalized. Do not change expectations, weaken assertions, substitute a locally reimplemented operation, or mark complete to avoid a blocker.