CI: cherry-pick shell:bash + OS-split test filename convention from PR 820

[eduralph]

Summary

Cherry-picks two companion CI fixes from feature/ci-cd-pipeline-upstream (= upstream PR 820) onto fork's maintenance/gramps60, so the fork-internal CI stops failing/hanging while we wait for 820 to land upstream:

• 28febdcd — shell: bash on unit-test-linux + integration-test steps
• 715e71dd — OS-split addon test filename convention (test_linux_* / test_windows_* / test_integration_*) + applying the convention to TMGimporter

Background

Issue 1: Bad substitution on Unit Tests (Linux) and Integration Tests

The container default shell on python:3.12-slim is /bin/sh (dash). The current workflow steps use bash-only parameter expansions:

mod="${f%.py}"
mod="${mod//\//.}"

Dash fails on the second line with Bad substitution, the script aborts, and unittest is never invoked. Visible in the April 19 maintenance/gramps60 run:

/__w/_temp/.../sh: 5: Bad substitution
Process completed with exit code 2

28febdcd adds shell: bash to the affected steps. Compile Check already sets it; Windows jobs inherit bash via job-level defaults.run.

Issue 2: Unit Tests (Windows) hangs for 6 hours

Visible in the same April 19 run: Unit Tests (Windows) in 6h5m1s — GitHub's job timeout. Cause: TMGimporter's tests/test_libtmg.py calls make_database("sqlite").load(":memory:", None), which deadlocks under conda-forge GTK + pip Gramps on Windows.

715e71dd introduces a filename convention (mirroring the existing test_integration_* convention) and applies it to TMGimporter:

Filename	Linux	Windows	Integration job
test_*.py	✓	✓
test_linux_*.py	✓
test_windows_*.py		✓
test_integration_*.py			✓

For TMGimporter:

• 7 pure-logic test classes (string codes, date parsing, repo-type / URL inference) stay in tests/test_libtmg.py — run on every OS.
• 13 DB-backed test classes (and the _Rec / _table / _make_db / _add_person / _MockUser helpers they use) move to tests/test_linux_libtmg.py — Linux only, no behavioural change.
• Windows job no longer touches the deadlocking imports.

Why fork-only

Both fixes are already in upstream PR 820 (commits 28febdcd and 715e71dd on feature/ci-cd-pipeline-upstream). When 820 merges and the fork resyncs, this cherry-pick becomes a no-op merge. No upstream PR is needed.

This PR exists only to clear the immediate red/hung CI on eduralph/addons-source maintenance/gramps60 until upstream catches up.

Verification

• yaml.safe_load parses the resulting workflow (7 jobs).
• All three ${var//pat/repl} sites verified under bash (two via step-level shell: bash, one via Windows-job-level defaults.run.shell: bash -el {0}).
• OS filter rules verified symmetric: Linux job skips test_windows_* + test_integration_*; Windows job skips test_linux_* + test_integration_*.
• Reproduced the original Bad substitution failure under dash and confirmed bash succeeds on the same expansion.
• TMGimporter: ran the testbed runner against the post-cherry-pick tree:

$ ./scripts/ubuntu/run-addon-unit.sh TMGimporter
Ran 175 tests in 8.640s
OK

Same total count as before the OS split (the 13 DB-backed classes now load via the dotted path TMGimporter.tests.test_linux_libtmg.* instead of …test_libtmg.*, with no test cases lost or added).

Remaining red on fork's maintenance/gramps60 after this lands

Out of scope for this PR (separate follow-ups; all continue-on-error: true so non-blocking):

• 4 addons missing po/template.pot → waiting on a gramps60 backport of upstream #831.
• 5 addons with plugin-registration errors (Collections Clipboard Gramplet, Query Quickview, lxml Gramplet, rebuild_types, TimePedigreeHTML) → real per-addon code bugs.
• WebSearch tests/test_filetable.py bare-import bug → addressed by #11 / upstream #833.
• Lint / Compile Check pre-existing failures → covered by 820's broader cleanup, not in this surgical cherry-pick.

🤖 Generated with Claude Code