Recover Codex profile policy and measure input preprocessing

[seathatflowsinourveins]

Scope

Repair missing-key recovery in existing Codex profile-v2 files and retain slot-bound Headroom/SocratiCode approval policy. Inventory token practice across native clients and SDK workers, and measure native gateway preview artifacts with promptfoo. RTK adoption remains PR #705; this PR removes its overlapping row/card changes.

Every preview/model/usage result is input-preprocessing evidence only. A3's native transformVars phase invokes the upstream preview CLI and direct inference; option 1's isolated integrated route/settings-hash/readback/retry evidence stays with the gateway owner. No integrated worker route or output-style acceptance claim.

• Base commit: f946c6d4ca988a17b6fa4392ecb488909f147883
• Lane: lane:foundation
• Owned paths: client renderer/template/tests; dated decisions and anti-pattern log; sanitized token-parity/gateway fixtures and evidence registration.
• Draft until the command center's cross-family read; cannot land before the v2026.10.05 re-pin. The command center owns the client-config templates, and the command center or user applies host configuration. No host apply from this lane in any window, including client-config apply, real-home RTK init or hook trust apply. No merge or shared gateway setting/combination/key changes.

SOTA sources

• openai/codex@a956835d… profile loader, native writer scope, MCP approval policy: reuse guarded additive merging for profiles; preserve operator choices and upstream tool annotations.
• DeusData/codebase-memory-mcp@v0.11.0 query annotations; Headroom@v0.37.0 local API; SocratiCode@v1.15.0 external readiness.
• native-agent-stack@4c897418f code-search plan:1073, owner decision:131-140: SocratiCode retrieval is expected pending configuration and the frozen semble/SocratiCode comparison, not a fault. No Qdrant/embedder/competing service installation or separate readiness handoff from this lane. The coordinator identifies semble as the working arm and assigns Claude D02 to fixwave-defects item 8.
• diegosouzapw/OmniRoute@23a11484… native preview, master-off/header boundary, native quality client usage limits: existing tools and request-local fixtures; no rebuilt harness or custom runner.
• promptfoo/promptfoo@34f74d34… native Responses passthrough, JSON rendering, native assertion context. Clean npm release installation; static fixtures and supported inline assertion expression.
• A3: promptfoo inline transforms, process shim, pre-repeat variable preparation; OmniRoute native preview CLI; Node24 execFileSync and createHash. Supported inline config calls upstream CLI; no custom runner/provider callback or credential reader.
• Worker pins and source paths: OpenHands default condenser, GPT Researcher selector, DeerFlow summarization default.

Evidence-class table

Fresh qualification: 48/48 exact final answers, exit 0; raw/RTK/Caveman each 16 calls. Native input totals 17950/13150/17758, output 1575/1684/1535. The earlier failed pilot is retained; all 96 benchmark calls have terminal native usage, while preparation/recovery and backend retries remain unknown. This does not establish complete-task savings or global settings.

A3's separately frozen native-transform phase: 48/48 exact answers, exit 0, 300947ms, 24 distinct pre-repeat previews. Input totals remain 17950/13150/17758; output 1646/1588/1516. All 144 distinct calls retain terminal usage; preview counts and cached/reasoning subsets are separate. Masked exported hashes remain masked, with retained-text verification derived against frozen constants. Inference headers report gateway3.8.51 and compression off, separately from installed canary source metadata. Input preprocessing only; unknown preparation/recovery/backend retries remain unknown.

Claim	Evidence class	Command / receipt
Profile merge preserves values/backups/guards and correctly reports failure	synthetic	targeted repository tests; token-parity receipt
Historical scratch Codex used RTK and selected MCP APIs	local_integration / native_proven	token-parity receipt; original pre-hook arguments absent; zero tiny-fixture savings
Current host profile applied	not established	historical guarded apply exited 1; no bytes changed; all future host application belongs to command center or user
Native gateway preview returned measured text and diagnostics	local_integration	frozen preview requests/metrics; preview estimates differ from provider tokens
Matched inference on frozen synthetic text	native_proven with synthetic oracles	retained failed pilot and separately frozen qualification, actual raw usage retained
Integrated worker/outputstyle or complete-task savings	not established	owner-isolated route/retries/preparation/recovery remain pending

Local commands run

python3 -m unittest -v tests.test_new_wsl_client_config tests.test_codex_hook_trust tests.test_token_stack_cards tests.test_native_token_ci
exit 0; 262 targeted tests (historical accepted implementation)

python3 -m unittest -v tests.test_adoption_docs_consistency
exit 0; 39 tests, one skip (historical implementation)

python3 tools/adoption/new_wsl_client_config.py --check
exit 0

promptfoo eval --config <frozen-pilot> --repeat 2 --max-concurrency 1 --no-cache --no-share --no-write --no-table --no-progress-bar
exit 100; 48 actual HTTP-200 calls, 29 exact passes/19 failures; retained

python3 scripts/validate.py
exit 0; 69 components, 10158 hashed files, 4 profiles, 202 receipts; integrity/scope only

python3 -m unittest -v tests.test_new_wsl_client_config tests.test_adoption_docs_consistency
exit 0; 217 tests, one skip, 70.913s

python3 scripts/validate_convergence.py --all-recorded --root . --json
exit 0; 32 records consistent, including this six-observation partial-usage experiment

python3 -m unittest -v tests.test_adoption_docs_consistency
exit 0; 39 tests, one skip, 15.017s after A1/A2 documentation correction

python3 scripts/validate.py
exit 0 after A1/A2 documentation/hash correction; same integrity counts

promptfoo eval --config evidence/artifacts/gateway-token-feature-evaluation-20261005/transform-vars/promptfoo.json --repeat 2 --max-concurrency 1 --no-cache --no-share --no-write --no-table --no-progress-bar --output <private-native-results.json>
exit 0; 48/48 exact answers; 24 unique preview invocations; 300947ms

python3 -m unittest -v tests.test_adoption_docs_consistency
exit 0 after A3 documentation changes; 39 tests, one skip, 16.070s

python3 scripts/validate_convergence.py --all-recorded --root . --json
exit 1 for duplicate attempt identifier in the expanded record; corrected to actual third experiment phase
exit 0; 32 records, nine gateway observations; consistency only

python3 scripts/validate.py
exit 0 after final A3 evidence registration; 69 components, 10163 hashed files, four profiles, 202 receipts

Decision record

• docs/decisions/2026-10-05-codex-token-parity.md
• docs/decisions/2026-10-05-gateway-token-features.md

Retain shared settings and the clean-default carrier holdout. Source capability, native preview, provider execution and complete-task acceptance remain distinct. Owner 99 triages hook gaps; gateway owner supplies isolated integrated evidence.

Host evidence

No files under evidence/hosts/ changed. Raw conversations and active configuration remain private; sanitized receipts retain hashes and failed conditions.

Checklist

• No workflow or hosting changes.
• No credential values/files read, printed or copied.
• No merge/main push; rootless/container boundaries preserved.
• Draft with exactly one lane label; evidence registry committed last.

[seathatflowsinourveins]

Folded into #770 by the PR triage of 2026-10-07 (the command center's ruling, item review-ns2604-coop-20261007T023012Z, section 3). Closed with a record, not merged. The co-op verified the fold at #770's head e4e1e60: 70/76 of this PR's files are byte-identical there, and the shared files (the registry, the client-config tool and its test, the decision records) are merged with this PR's added lines present except where the merge re-hashed or re-counted them; #770's description names this fold, its composition record (docs/decisions/2026-10-07-codex-profile-environment-content-fold.md) and the donor-byte ledger under evidence/artifacts/codex-profile-env-fold-20261007/. This branch codex/ns2604-codex-token-parity-20261005 stays on origin at b63381f. Reopen trigger: none, because its content lives in #770.