test: prune the dead tests, and give the suite a shape

[hugolytics]

The problem

The suite was 359 files and 3,192 tests, green in 33 seconds. Speed was never
the issue — sprawl was. tests/unit/ was 269 files in one flat directory, and
some of them were guarding code that nothing calls any more.

Three things were wrong, and only the first is obvious:

1. Tests outliving their subject. Six modules test code with no caller
   anywhere in src/ or scripts/ — superseded when haunting moved to
   fateforger/haunt/ and the flow moved to the stage-gated agent. One of them
   asserts the line numbers of methods in a code-navigation helper.
2. A refactor scaffold that outlived its refactor. _parse_event_dt exists
   at three call sites, and each grew its own branch-by-branch table written,
   per its own docstring, "before and after the refactoring to Pydantic
   dispatch". The refactor landed. All three now delegate to one
   EventDateTime.to_datetime, whose table is tested directly elsewhere.
3. Doubles copied instead of shared. Nine modules had each grown their own
   DummyClient for the Slack web client, differing only in the timestamps
   they invent. Where a double was shared, it was shared by importing another
   test module — which works only while both happen to be collected.

What changed

	before	after
test files	359	341
tests/unit/ flat files	269	0
tests	3,192	3,137 (+89 slow, unchanged)
lines	76,892	75,820

• Pruned six dead-subject modules (44 tests, 6 of them already skipped).
• Collapsed the three duplicated date-parsing tables. What each call site
  still owns — the ISO-string branch — stays, because those genuinely differ:
  reconcile keeps the parsed offset, the planner converts into tz. That
  divergence now has a test naming it, which it did not before.
• Restructured tests/unit/ into subpackages beside the tmbx/ one that
  already worked this way: timeboxing/ (96), slack/ (49), constraints/
  (34), core/ (33), tmbx/ (22), haunt/ (14), tasks/ (9), schedular/ (7).
• Extracted tests/doubles/ — one Slack client recorder replacing nine, plus
  the reconciler, planning, required-block and kernel doubles that modules were
  importing out of each other.
• Added tests/repo.py. Twenty files found the repo root by counting
  parents[N], so moving a test one level down broke it with a
  FileNotFoundError that reads like a missing fixture. It walks up to
  pyproject.toml instead.
• Folded 20 one- and two-test fragment files into nine subject modules.

Proof

Every lost line is accounted for. Coverage measured per line, before and
after, on the same command:

before  22,221 / 28,593  77.71%
after   21,947 / 28,593  76.76%

All 274 lost lines are in the six modules with no caller, plus the
llm/factory and tools_config branches only their tests reached. No other
file lost a covered line.

The one exception, timeboxing_session_store.py:157, is the raise StaleSessionRevision inside a concurrent-save race. Per-test coverage contexts
run against origin/main show that line is not covered reliably there
either — the concurrent-save test wins the race nondeterministically. Not a
regression from this branch.

Merges cannot silently shadow. The merge tool refuses any merge where two
sources define the same top-level name.

Three near-duplicates were investigated and kept. planning_day_for against
the starter that calls it, two ConstraintStore prune scenarios, two MCP
clients' empty-loader guard — same test name, different seam. Only genuine
duplicates were removed.

3,137 passed, 4 skipped, 89 deselected, 1 xfailed
3,231 collected including the slow evals (not run — they hit OpenRouter)

What I did not do, and why

The dead source modules are still here. This PR is test-only. Deleting them
is a source change and wants its own PR. The evidence, if you want it:

admonisher/{base,calendar,commitment}, schedular/diffing_agent,
timeboxing/{flow,prompts,state,notebook_entrypoints},
slack_bot/{relay_agent,topics}, tools_config/ — no caller in src/ or
scripts/, last touched January–February 2026, and now visibly at 0% coverage.

The docs are in this PR, not a follow-up. tests/README.md was a
hand-maintained per-file index — which is why it rotted, and why it still
listed test_timeboxing_flow.py as live logic. It now describes the shape
(subpackages, tests/doubles/, tests/repo.py) instead of the inventory.
CALENDAR_QUERY_LOCATIONS.md and MIGRATION_ARCHIVE_TO_CALENDAR_HAUNTER.md
carry superseded banners pointing at fateforger/haunt/, with their original
text intact as a record.

Two claims in that rewrite were wrong and are fixed in a follow-up commit: it
said the fast suite is "what CI and pre-merge checks run" (nothing in
.github/workflows/ runs pytest at all), and it told you to run
.venv/bin/python while explaining why worktrees break the venv — a worktree
has no .venv, so that path resolves to nothing exactly where the warning
applies.

timeboxing/ is still 96 files. It is the largest genuine subject in the
repo and folding it further needs judgement per file, not another pass of the
merge tool.

One gap this surfaced but did not fix: fateforger/haunt/orchestrator.py
has its own capped exponential backoff and no test reaches it. The capped-backoff
test I deleted belonged to the dead BaseHaunter, so nothing regressed — but
the live one has never been covered. Worth a ticket.

Checklist for review

• The six pruned modules really are dead — spot-check one with
  grep -rn CalendarHaunter src/ scripts/
• The kept _parse_event_dt string-branch tests are the right call: the two
  implementations disagree about timezone conversion, and neither is
  obviously wrong
• The subpackage boundaries match how you think about the system —
  constraints/ split out of timeboxing/ is the one most open to argument
• tests/doubles/ is the right home, and one Slack recorder for nine call
  sites does not over-generalise
• Whether the dead source modules should go in a follow-up or be left

🤖 Generated with Claude Code