test: pin apply_multi_node_invalid_variants CONC filter (PRPUNDIT-10)

[jiagaoxiang]

Summary

• Locks the single-node no-op, multi-node --cuda-graph-max-bs N < CONC drop/keep split, CONC=0 guard, and unparseable CONC fallback to 64.

Closes test gap PRPUNDIT-10.

Test plan

• PYTHONPATH=src pytest src/hyperloom/inference_optimizer/tests/test_grid_variant_filter.py

[zoroyihan7]

Thanks — good coverage of the observable behaviour: both flag spellings, the at-threshold boundary, order preservation, and kept is grid on the no-op path.

Mutation testing found three producer-side holes, all in the same area.

1. The regex is applied with .search() but every test puts the flag first. All four tests build extra_server_args as a single flag that starts the string ("--cuda-graph-max-bs 32", "--cuda_graph_max_bs=8", "--chunked-prefill-size 8192"). Mutating .search( → .match( at _grid_variant_filter.py:201 leaves all 4 green. Real variants are multi-flag strings (see _grid_server_args.py:369-372 and the explore.py:1036 call site), so a regex-anchoring regression would ship silently. One variant with the flag in non-leading position — e.g. "--tp 8 --cuda-graph-max-bs 32" — closes it.

2. The 64 default has three producers and only one is pinned. Line 197 is int(os.environ.get("CONC", "64") or 64):

producer	pinned?
ValueError → 64	✅ by test_unparseable_conc_falls_back_to_64
get default "64" (CONC unset)	❌ mutating "64" → "1" stays green
or 64 (CONC empty string)	❌ mutating or 64 → or 1 stays green

Every test does monkeypatch.setenv("CONC", ...), so CONC-unset and CONC-empty are never exercised. Two delenv/setenv("CONC","") cases would cover both. The or "" None guard on line 201 is likewise unpinned.

3. test_conc_zero_does_not_drop can't fail. Deleting conc > 0 and from line 202 leaves all 4 green — and it can't be otherwise with the code as written, since the regex captures (\d+) so the parsed value is always >= 0, and n < conc is already False for every conc <= 0. The guard is genuinely unreachable as a differentiator. Either drop the redundant guard from production and keep this test as a behaviour doc, or add a comment saying the test pins the observable behaviour rather than the guard line. (Not asking you to change production here — just flagging that this one doesn't buy what the ticket assumed it would.)

Nits:

• kept is grid (identity) is stricter than the docstring and stricter than the sibling filters — apply_aiter_moe_pin_filter:269 and apply_user_skip_list:577 both return list(grid) on their no-op path. A future harmless defensive copy would turn this red. If the identity is deliberate, a one-line comment would help; otherwise == grid plus an order assertion is safer.
• The dropped reason is asserted only via the "CONC=64" substring, which is shared by the explicit-CONC and fallback tests, so it doesn't distinguish them on its own. Asserting the cuda_graph_max_bs=32 half too would make the row self-describing.

[jiagaoxiang]

Addressed the three mutation holes and dropped the unreachable conc > 0 guard (regex (\d+) already guarantees conc >= 0; n < conc is false for every n >= 1 when conc == 0, so behaviour is unchanged). Also pinned .search vs .match with a non-leading flag, CONC-unset / CONC-empty / extra_server_args is None, and the cuda_graph_max_bs=32 half of the dropped reason. Left kept is grid as identity with a comment — that is the production single-node return grid, [] contract.

[zoroyihan7]

Re-reviewed. All three holes closed, and dropping the conc > 0 guard with the reasoning recorded in the test comment is the right call.

Five mutations, all red now:

• .search( → .match( — red (the non-leading-flag variant does it)
• os.environ.get("CONC", "64") → "1" — red
• or 64 → or 1 — red
• v.extra_server_args or "" → v.extra_server_args — red
• < conc → <= conc — red (4 tests)

I also checked the guard removal is behaviour-preserving: the regex captures (\d+), so the parsed value is always >= 0 and n < conc is already False for any conc <= 0, negatives included. 8 passed. LGTM.

One heads-up: this PR now touches _grid_variant_filter.py, so it isn't test-only any more. I'd added skip-e2e-test when it was — removing it, since the workflow scopes that label to PRs that don't touch runtime code.