test(executor): add unit tests for wrap_executor and its helpers

[pmrv]

Motivation

A branch-coverage pass over tests/unit (pytest-cov, --cov-branch)
identified src/fleche/executor.py as the least covered module in the
package: 16.0% of statements, all 12 branches missed. Every other
src/fleche/* module was above 80%, most above 93%.

  16.0%    38 stmts    30 missing  br: 12/ 12  src/fleche/executor.py
  81.5%    21 stmts     3 missing  br:  6/  2  src/fleche/__main__.py
  83.3%    12 stmts     2 missing  br:  0/  0  src/fleche/_attrs.py
  86.7%    13 stmts     2 missing  br:  2/  0  src/fleche/__init__.py
  91.4%   369 stmts    26 missing  br: 82/ 13  src/fleche/remote.py
  ... (rest ≥ 91%) ...

The gap comes from wrap_executor and its two helpers
(_is_fleche_function, _split_submit_kwargs) only being reached
end-to-end through the process-pool tests in
tests/integration/test_parallel_execution.py — the module itself had
no unit-level test file. That makes the executor branches a soft spot in
regression coverage: each subprocess spin-up is slow, and none of the
branches (already-wrapped shortcut; signature-parse fallback;
cache-hit-skips-submit) reads like something the integration suite is
guaranteed to keep exercising as the surrounding code evolves.

What this changes

Adds tests/unit/test_executor.py, a 15-test file that drives the
module in isolation with three lightweight executor stubs:

• _StubExecutor — records what gets submitted; runs the callable
  synchronously in a completed Future.
• _NeverSubmitExecutor — raises on any submit call, so a
  cache-hit branch that fails to short-circuit fails loudly.
• _KeywordOnlyExecutor — declares a keyword-only resources=
  param, modelling the HPC-executor shape (executorlib / dask) that
  motivated _split_submit_kwargs.

Test groups:

Class	Behaviour pinned
TestIsFlecheFunction	plain callable, fleche-decorated callable, and non-SimpleNamespace .fleche attribute (branch coverage for isinstance check)
TestSplitSubmitKwargs	no-keyword-only pass-through, keyword-only split, empty-kwargs edge case, inspect.signature TypeError fallback (via unittest.mock.patch)
TestWrapExecutor	returns same instance, marks submit._fleche_wrapped, idempotent second wrap, non-fleche BoundWrapper.bind, already-BoundWrapper passthrough (PR #724), cache-hit skips submit, cache-miss binds and forwards, keyword-only kwargs partitioned correctly

Coverage impact

Measured with pytest tests/unit --cov=fleche --cov-branch:

• src/fleche/executor.py in isolation goes from 16% / 0% branch
  to 100% / 100% branch (0 missing lines, 0 missing branches) — no
  other module's coverage regresses.
• Overall tests/unit branch coverage was 94.0% before; the 30 new
  statements + 12 new branches lift it further, with no offsetting
  drop elsewhere.

Test plan

• pytest tests/unit/test_executor.py -v — all 15 pass locally.
• pytest tests/unit/test_executor.py --cov=fleche shows
  src/fleche/executor.py at 100% statements + 100% branches.
• The three new stub classes don't depend on any optional extra
  (no cloudpickle / no executorlib / no ssh path) — the file runs
  on the minimum-deps CI matrix.

---

Generated by Claude Code

[claude]

🤖 CI failure summary for da7098c on claude/gifted-meitner-kt6uww

Summary
Three executor tests are failing across Python 3.11, 3.12, and 3.13. The tests expect BoundWrapper instances to be submitted to the executor, but wrapped functions are being double-wrapped instead.

Root cause
The wrap_executor function is wrapping callables that are already BoundWrapper instances. The tests show:

• test_non_fleche_callable_is_bound_before_submission: expects isinstance(submitted_func, BoundWrapper) but gets a nested wrapper
• test_already_bound_wrapper_is_forwarded_as_is: expects the same BoundWrapper object but gets a new one wrapping the original
• test_cache_miss_binds_and_submits: similar double-wrapping issue

AssertionError: assert False
 +  where False = isinstance(BoundWrapper(...), BoundWrapper)

The problem is that wrap_executor should skip wrapping if the callable is already a BoundWrapper, but it's currently wrapping it again.

Suggested fix
In the wrap_executor function (likely in src/fleche/executor.py), add an isinstance check before wrapping:

• Check if the callable is already a BoundWrapper
• If yes, pass it through unchanged
• If no, proceed with binding as normal

This ensures idempotence so that pre-bound wrappers aren't double-wrapped.

---

Generated by Claude Haiku 4.5.

[pmrv]

@claude ^

[claude]

Claude finished @pmrv's task in 6m 48s —— View job

---

Root cause found — fixed a test-isolation bug, not wrap_executor itself

• Gather context (PR diff, prior CI-failure comment)
• Reproduce/confirm the failure (read the actual CI job logs, not just the bot's summary)
• Determine root cause
• Implement fix
• Push

The Haiku bot's diagnosis was wrong. src/fleche/executor.py already has the isinstance(func, state.BoundWrapper) short-circuit (added in #724, well before this branch was cut) — it's untouched by this PR, which only adds tests/unit/test_executor.py.

Reading the actual job log (not just the summary) shows the real signature of the bug:

E       AssertionError: assert False
E        +  where False = isinstance(BoundWrapper(func=<function ...plain...>, ...), BoundWrapper)

isinstance returning False against an object whose repr() is literally BoundWrapper(...) is the classic symptom of two distinct class objects with the same name — not a double-wrap bug (the apparent nested-BoundWrapper repr in the second failure is a downstream effect of that, not the cause).

Tracing it down: tests/unit/config/test_config.py has a restore_fleche_state fixture (used by 3 tests) that calls importlib.reload(fleche.state) to simulate fresh-process config loading. That gives fleche.state a brand-new BoundWrapper class object. But fleche/__init__.py does from .state import BoundWrapper at package-import time, so fleche.BoundWrapper keeps pointing at the pre-reload class forever after — while src/fleche/executor.py does from . import state and looks up state.BoundWrapper fresh on every call, so it sees the new class. Since tests/unit/config/ sorts before tests/unit/test_executor.py alphabetically, every full-suite run silently diverges the two references before the executor tests even start, breaking any isinstance check that spans the public API (fleche.BoundWrapper) and the internal one (executor.py's state.BoundWrapper).

Fix (tests/unit/config/test_config.py): the restore_fleche_state fixture now also reloads the fleche package itself after reloading fleche.state, resyncing fleche.BoundWrapper (and the other re-exports) with whatever fleche.state currently holds. fleche/__init__.py has no risky side effects to re-run (just rebinding names), so this is safe. Pushed as a793292.

Note: I wasn't able to run pytest locally in this sandbox (Python execution requires interactive approval this environment doesn't grant), so this fix is verified by log analysis and code tracing rather than a local run — worth confirming the new CI run goes green. If you'd like me to also run the suite here, granting python/pytest execution in --allowedTools would let me do that directly next time.
·

[pmrv]

@claude still broken

[claude]

Claude finished @pmrv's task in 4m 0s —— View job

---

Fixed — real root cause this time

• Gather context (PR diff, prior fix, CI-failure history)
• Get latest CI run logs
• Determine root cause
• Implement fix
• Push

Why the previous fix (a793292) didn't hold: it reloaded fleche/__init__.py after reloading fleche.state, to resync fleche.BoundWrapper with the brand-new BoundWrapper class object that importlib.reload(fleche.state) creates. But tests/unit/test_executor.py:21 does from fleche import BoundWrapper at module-collection time — before any test body runs. Resyncing fleche.BoundWrapper later doesn't touch that already-bound name; it stays frozen to the original (pre-reload) class forever. Once the config tests reload fleche.state later in the run, executor.py's live state.BoundWrapper lookup (src/fleche/executor.py:92) sees the new class while test_executor.py's captured name still points to the old one → same three isinstance failures as before, byte-for-byte.

Actual fix (tests/unit/config/test_config.py): stop letting a second BoundWrapper class escape at all. _reload_fleche_state() now reloads fleche.state and then restores fleche.state.BoundWrapper = original_bound_wrapper. Because importlib.reload re-executes a module inside its existing __dict__, the restored class's bind() classmethod still resolves _CACHE/_METADATA against the freshly reloaded ContextVars — only the class identity is preserved, not its stale config. Every reference taken anywhere (before or after the reload — this file, test_executor.py, tests/integration/test_parallel_execution.py, user code) now points at the same class, so no downstream isinstance check can ever diverge. This is more robust than the previous approach, which only patched the one path (fleche.__init__) that happened to be involved in this specific failure.

Verification note: as with the previous pass, I could not execute pytest/python directly in this sandbox (execution requires interactive approval this environment doesn't grant) — the fix is verified by tracing importlib.reload semantics and the exact import order in the failing log, not a local test run. Worth confirming the new CI run goes green; granting python/pytest in --allowedTools would let me verify directly next time.
· branch