fix(stand): date-independent journeys + seed; e2e-stand runs on PRs from the checkout

[ktursunov]

What

The two evidence drill-down journeys (test_git_commit_bucket_opens_and_exports_supporting_data, test_table_total_opens_supporting_data_for_the_whole_period) fail on some calendar dates and pass on others — they are red in the current nightlies and in dispatch run 31560385144.

Why

Both page objects click the first body row of a timeseries table. The first row of a trailing-month window is the leading partial week bucket, whose overlap with the window can be a single day. A bucket with no activity renders —, and MetricTableValue only wraps a valued cell in a <button> — so on those dates there is nothing to click and the locator times out after 30 s. The seeded activity is weekday-shaped and Poisson-drawn, so whether the leading bucket is drillable depends on which day the window starts (weekend-only overlap being the common empty case).

Fix

• Page objects target the first drillable cell instead of the first row: open_bucket_evidence / open_total_row_evidence click the row group's first button; open_first_commit_bucket filters to the first row whose Commits cell carries a button, staying pinned to that column so the dialog opened is the one the test names.

• Git seed floors commits at one per person-day (weekday_multiplier still shapes volume), so every bucket of the git surface is drillable whatever day the stand is seeded. PROFILE.md regenerated for the new seed_revision.

• Seed link parity (review follow-up): the commit floor had broken the RNG re-derivation contract — seed_class_git_pull_requests_commits still computed the day's commit count with the raw Poisson draw, so a zero-draw day carried a floored commit with no PR-link rows, and a short day (fewer commits than PRs) dealt empty slices to the tail PRs; either way the PR loses its author attribution in gold's pr_commit_emails join. Both generators now share one _daily_commit_count helper and the round-robin deal wraps on short days. tests/test_git_links.py locks the parity at the emitted-row level, with the anchor pinned so the fixture window provably contains a zero-draw PR day.

CI workflow (absorbed from #2458, then extended)

e2e-stand.yml is converted from a merge-queue-only gate into a per-PR one, and the stand's scope is broadened to any change under tests/**, src/backend/**, or src/frontend/**:

• Runs on PRs for real. The event guard that skipped every job on pull_request is gone: a PR whose diff matches the new paths: filter (mirroring the changes job's regex — the two lists are documented as kept in sync) runs api-smoke and ui-journeys, with superseded runs cancelled by the concurrency group. merge_group (no paths: support), the nightly, and dispatch keep working via the changes job. While the paths filter exists, Stand E2E can only be made required on merge_group, never at PR level — a required check whose workflow never triggers reports as pending forever; the current main ruleset requires neither, so nothing breaks today.
• Backend diffs build the backend. test-stand up refuses a tree that changes src/backend/ rather than testing it against main's pulled images, so changes reports a backend flag and the lanes translate it into up --build-backend plus a wider timeout.
• ui-journeys runs host-side from the checkout on every event, so the lane always tests the ref's own test code. It previously ran the suite baked into the published insight-ui-tests:latest image, which only moves on a main push — a PR lane therefore ran main's old test code against the PR's stand and could not go green on the very PR fixing it. The host's localhost:<port> satisfies the same __Host- cookie origin constraint the image's netns join existed for.
• The ui-tests image is removed entirely (Dockerfile, build/publish jobs, change filters, tests/versions.yaml sync ledger): after checkout mode, its only consumer was the pipeline validating the image itself. test-stand test --image <ref> remains as a mechanism for a locally built runner; insight-ui-tests stays in ghcr-cleanup so published versions keep getting pruned.
• The seed package's new tests are pytest-style; build-images.yml's seed validation installs pytest transiently into the throwaway container, so the published seed image still ships without it.

Verification

Against a freshly re-seeded compose stand: both previously failing journeys pass, and the full tests/stand suite is green — 296 passed, 3 xfailed, 2 xpassed (the xpasses are the known #2361 flake entries).

This branch's UI-test code was also run host-side against a live stand seeded without the commit floor on a date where CI reproduced the flake — 17/17 journeys pass, proving the locator half absorbs the undrillable leading bucket on its own. And the converted workflow demonstrates itself on this PR: the PR-triggered e2e-stand runs execute both lanes from this branch's checkout, 17/17 ui-journeys, Stand E2E green.

Summary by CodeRabbit

• Bug Fixes

  • Improved activity generation so eligible people receive at least one commit per day while preserving realistic volume limits.
  • Improved evidence navigation in group and person views by reliably opening the first available drill-down option, including total rows and partial-week data.

• Documentation

  • Updated the documented seed revision to reflect the latest generated dataset.

[coderabbitai]

📝 Walkthrough

Walkthrough

The seed generator now creates at least one commit per eligible person per day. UI page objects now locate drillable evidence controls across partial or empty leading buckets and footer rows.

Changes

Seed generation and evidence navigation

Layer / File(s)	Summary
Daily commit generation src/ingestion/tools/seed/PROFILE.md, src/ingestion/tools/seed/insight_seed/generators/git.py	The Git seed generator clamps daily commit counts to at least one while retaining Poisson volume and COMMITS_CAP. The documented seed revision is updated.
Evidence navigation tests/stand/ui/pages/group_dialog.py, tests/stand/ui/pages/person_view.py	Evidence navigation searches body and footer rowgroups for available buttons. Commit bucket navigation skips leading rows without a drillable Commits button.

Estimated code review effort: 2 (Simple) | ~15 minutes

Possibly related PRs

• constructorfabric/insight#2253: Modifies the relocated seed tooling and Git seed generator.
• constructorfabric/insight#2255: Updates UI metric evidence and drilldown navigation.

Suggested reviewers: cyberantonz, aleksdotbar

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Title check	⚠️ Warning	The title accurately describes the date-independent journey and seed fixes, but it also claims PR checkout execution changes that are not present in the changeset.	Remove the unsupported “e2e-stand runs on PRs from the checkout” clause or include the corresponding workflow changes.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.

✨ Finishing Touches 💡 1

🛠️ Fix failing CI checks 💡

• Create stacked PR
• Commit on current branch

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/stand-date-independent

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@src/ingestion/tools/seed/insight_seed/generators/git.py`:
- Around line 91-97: Apply the same minimum-one commit-count rule used by
seed_class_git_commits to seed_class_git_pull_requests_commits, preferably by
reusing a shared helper, so zero Poisson draws still generate the commit index 0
and its PR link. Add a regression test covering a zero Poisson draw and
confirming the commit and link are created.

🪄 Autofix

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: fc461118-8d01-4b5c-be9e-ab18b9ca8595

📥 Commits

Reviewing files that changed from the base of the PR and between 825061d and e76ef7e.

📒 Files selected for processing (4)

• src/ingestion/tools/seed/PROFILE.md
• src/ingestion/tools/seed/insight_seed/generators/git.py
• tests/stand/ui/pages/group_dialog.py
• tests/stand/ui/pages/person_view.py

[ktursunov]

Dispatch run on this branch: 31567069878 — api-smoke ✅, ui-journeys 16/17 with test_table_total_opens_supporting_data_for_the_whole_period still red. That residual failure is expected and does not indicate the fix is wrong:

The ui-journeys lane runs the suite baked into the published insight-ui-tests:latest image ("pulled, never built" — see the workflow comment), and :latest is tagged only on a main push. So the lane exercised main's page objects, not this branch's — the failing call log shows the old row.first locator. Two confirmations inside the same run:

• The git journey went green in CI for the first time in days with the old locator still in place — that is the seed floor working: the leading bucket's Commits cell is now always drillable, so even a first-row click lands.
• The table-total journey ran the old locator against the task-delivery table, whose sparse issue-closes legitimately leave the leading partial bucket — (no floor there, deliberately — close volume is realism the seed keeps).

The page-object fix is proven on a live stand by a host-side run of this branch's code: both previously failing journeys pass and the full tests/stand suite is green (296 passed / 3 xfailed / 2 xpassed). Once this merges, build-images.yml republishes the ui-tests image from main and the lane picks the fixed locators up.