[amar-python]

Summary

Tests only. No product code changes. Four files.

This PR started as a larger batch (#167, #168, #191–#193). While it sat in draft, main moved 70 commits and absorbed most of it: #167 (test layout, #221), #168 (accessibility job, #222) and #195 (#213), on top of #169, #193 and #194 from #202. I merged main in and dropped everything it already covers, so what is left is only what main still lacks. The earlier description of this PR (layout, accessibility job, a registry-race loader fix) no longer applies.

Issue	What this PR adds
#191	A real concurrency test for the registry, replacing the existing one
#192	Checks that each ragged row's reason is stored, and the same accounting for T&E mode
#193	Assertions that the XLSX cell is stored as text with no formula

#167, #168, #169, #194, #195 need nothing from this PR.

Changes

#191, tests/system/test_registry_race.py. The existing test used no barrier and accepted status: "error" as a valid loser outcome, so it could not fail for the race it was written about. This one releases uploads together with a barrier and repeats 40 rounds, for three cases, because the guard differs:

• identical file, same name: the per-dataset advisory lock serialises them; exactly one wins, the rest get an attributable duplicate_file
• identical content under different names: different datasets, so different locks; only UNIQUE(file_hash) stands in the way
• same name, different content: since M5 this is the next version, not a rejection, so it asserts distinct, gapless versions

#192, tests/system/test_ragged_rows.py. The existing tests check the accounting identity and a reason per ragged row in the response. These add that every reason is also stored (GET /api/csv/uploads/{id}), and the same checks for T&E mode, with a file mixing good rows, an in-file duplicate and several ragged rows.

#193, export.test.ts. The existing XLSX tests read values back through sheet_to_json. The issue asks for the cell to be asserted as stored as text, the property that stops a spreadsheet evaluating it: added for =cmd, +cmd, -cmd, @cmd and a tab-prefixed value, in the error and audit reports.

Evidence

PostgreSQL 16, all four environments deployed, on main (018edc2) plus this PR.

• Backend: 635 collected, 635 passed, 0 failed, 0 skipped. Frontend: 180 passed. Lint, bandit and the catalogue check pass
• Unmodified main: the race tests pass 12 of 12 runs
• Mutation checks, each fails when its guard is removed:
  • UNIQUE(file_hash) dropped: the different-names race test fails 6 of 6
  • per-dataset advisory lock removed: both same-name race tests fail 4 of 4
  • extra_cells() returning []: 6 of 9 tests in test_ragged_rows.py fail
  • sanitizeForXlsx guard disabled: 11 of 41 tests in export.test.ts fail

Correction to my earlier claim

I previously reported that the #191 test found a real defect, where a losing upload could get a generic "database error" instead of a duplicate rejection, and I proposed a loader fix. That was true on the old main. It no longer reproduces: M5 changed table naming and added the dataset lock, and the test passes 12 of 12 without any fix. I dropped the fix rather than ship a change for a defect that is gone. The same race still exists in principle if the lock is removed, which is why the lock-removal mutation above fails the tests.

Not verified

• CI has not run on the latest head (7b6beb7).
• Microsoft Excel was not available. A LibreOffice 24.2 check of the exported XLSX (done earlier) showed the guarded cells rendering literally, not evaluated.
• Still open from the original batch: the product owner's confirmation of the Gate: measure api/services coverage with floors for the three loaders #194 coverage floors (79/84/96).

Refs #191, #192, #193.

https://claude.ai/code/session_011T6Wqj8HS5EUEcvTUM24SL

---

Generated by Claude Code