[structural-drift-reviewer-codex-ladder] Delegate the client reviewer effort ladder to the server's codex model gate #6817

Slice audited: client-side hand mirrors of server sources of truth (client/src/lib, client/src/utils, client/src/components/*) — structural-drift audit, 2026-09-10.

Problem

• client/src/lib/reviewerPins.js:113-126 inlines a copy of codexEffortLevelsForModel (server/lib/providerModels.js:155-162) with its own regex and a hardcoded ultra-model list ['gpt-5.6', 'gpt-5.6-sol', 'gpt-5.6-terra', 'gpt-6-astra'] (the server's CODEX_ULTRA_MODELS, line 137). The copy landed in 255032a (Reviewer effort pins still offer codex 'minimal' for gpt-6 models that reject it #6447) after the server list had already moved twice without the client (e8efea4 introduced three ultra ids; f4df117 added gpt-6-astra).

• The copy is wrong today, not merely at risk. Running both sides on the same inputs (server reviewerEffortLevels('codex', model) from reviewerConfig.js:595 vs the client function):

  model	server	client
  gpt-5.6, gpt-5.6-sol, gpt-5.6-terra	minimal … max, ultra	minimal … max
  gpt-6-astra	low … max, ultra	low … max, ultra
  gpt-6, gpt-5.5, o3, null	equal	equal

  The client consults its ultra list only inside the gpt-6+ branch, so the gpt-5.6 family never gets ultra.

• The drift gate cannot see it: server/lib/reviewerConfig.test.js:396-405 ("client mirror of the reviewer vocabulary") compares reviewerEffortLevels(reviewer) with no model, so the model-gated branch is never exercised. The client file's header (reviewerPins.js:19) cites a "client mirror of the reviewer effort ladders" test that no longer exists (it went with providerModels.mirror.test.js in c82dcdf).

Impact

A codex reviewer pinned to a gpt-5.6 model cannot be given the ultra effort in ReviewerPicker although the server (and the CLI) accept it — the exact "hide a tier their CLI accepts" failure the drift test's own comment describes. The next change to CODEX_ULTRA_MODELS or CODEX_NO_MINIMAL_MODEL_RE repeats #6447 in the other direction (offering minimal to a model that answers HTTP 400).

Fix (decided)

1. server/lib/providerModels.js: export codexEffortLevelsForModel.
2. client/src/lib/reviewerPins.js#reviewerEffortLevels: replace the inline branch with if (slug === 'codex' && model) return codexEffortLevelsForModel(model);, imported from '../../../server/lib/providerModels.js' (the leaf client/src/utils/providerModels.js already imports from; reviewerPins.js stays dependency-free so the server suite can keep loading it). Delete the local ultra list and regex. Fix the header comment to name the real test.
3. Extend "matches server reviewerEffortLevels for every reviewer" in reviewerConfig.test.js with a codex model matrix — gpt-5.6, gpt-5.6-sol, gpt-6-astra, gpt-6, gpt-5.5, o3 — asserting toEqual against the server for each. Confirm the test is red against the current client file before applying steps 1-2.

Acceptance criteria

• grep -n "gpt-5.6\|gpt-\[6-9\]" client/src/lib/reviewerPins.js returns nothing.
• reviewerConfig.test.js compares codex ladders for at least the six model ids above and fails on the pre-fix client file.
• In ReviewerPicker a codex reviewer with model gpt-5.6 offers ultra.
• Server and client suites green.

Scope: small

Dispatch: model:light — a well-specified two-file edit plus a test matrix. effort:medium — the parity matrix has to cover both ladder gates (ultra list and no-minimal regex) and be proven red first.