Use the vendor TypeSafe SDK for crosswalk transport and retry - #725

Scope

The crosswalk's call() duplicates HTTP transport and retries already shipped by TypeSafe. Replace that loop with the published typesafe-sdk==0.7.2 client and its native retry policy, imported only for judge. Preserve request construction, validation, thresholds, ledger code and recorded judgments.

• Base commit after the requested rebase: 4af7417b7d4db8936c0136235a86717f31d16f27 (descends from 77d7516d8).
• Lane: lane:foundation.
• Owned paths: tools/sota-convergence/gap_crosswalk.py, its new SDK test, dated decision, compact synthetic evidence and the evidence registry.
• Claude cross-family read completed: 0 P1, 0 P2; both P3 documentation corrections are addressed. The command center authorized leaving draft after these fixes; the command center alone lands the PR. This is separate from the currency handoff in Withdrawn NativeStack2604 currency qualification handoff #721.
• Rebased conflict-free onto current main on 2026-10-05. The five owned source/evidence files remain byte-identical to their original qualification; dated evidence retains its original input revision. Main's registry was restored and native registration rerun, preserving every main entry and all non-file registry fields.

SOTA sources

• typesafe-ai/[email protected], f078f1e2: native client, structured Choice types and raw-question normalization. All three prerequisite compatibility gates pass at this pin.
• Usage, choice probabilities and public HTTP metadata preserve the existing validator input and optional request ID.
• The strict SDK schema uses extra="ignore" and strict=True, requires confidence, and drops unknown answer types before validate()/ledger. No modeled field is lost on any of the 695 retained successful rows (2,254 choice answers); future schema-mismatched responses can change acceptance or lose extra fields. The decision now makes this boundary explicit.
• RetryPolicy and native retry-count header. Explicit settings retain five attempts, 429/529 only, 1/2/4/8-second delays, no jitter or Retry-After, no connection/timeout retries and a 60-second attempt timeout.
• Vendor-owned headers explain the authorized User-Agent change to typesafe-sdk/0.7.2. pydantic/[email protected], d426e2c4: public elapsed contract supplies successful-attempt HTTP timing, excluding SDK decoding; the prior timer also included final JSON decoding. Earlier attempts/backoff remain excluded.
• Upstream MockTransport seam supplies the offline test boundary. PyPI 0.7.2 publishes the wheel; its downloaded SHA256 matches 0a961148187d52e18276ed7f2d02617cfac48e3b97673cb631a8749397d43d1e. Installation used an isolated uv environment with source builds disabled.
• Repository hot-file protocol at 4af7417b: native host_receipts.register_file updated the five owned files from main's registry and re-registered the revised decision; the registry is committed last at ad6979a866317456cf80e32855c0a591f8371d88.

Evidence-class table

The stub tests are locally authored integration checks through the unchanged SDK implementation, not an unchanged upstream suite or live provider acceptance. No live judge call occurred. The original environment-only credential consumer remains unchanged. No credential inventory, install script, host receipt or host-stack pin changed.

Dependency pinning in the runtime install hint remains an explicit limitation: the hint pins only typesafe-sdk==0.7.2. The observed dependency set includes httpx2==2.13.1, pydantic==2.13.5 and tenacity==9.1.4; another resolution can select versions outside that tested set. The optional hint change is deferred, as allowed by the cross-family verdict.

Local commands run

Checks used private scratch and an isolated uv cache; the private environment path prefixes are recorded in the lane status. Runtime checks are reused from the unchanged code/test inputs at 9a87ae07; the P3 follow-up changes only the decision and its registration. Validation was rerun before the doc/registry commits and after the final registry commit.

python3 tools/sota-convergence/gap_crosswalk.py build --check
exit 0: crosswalk up to date

python3 -S tools/sota-convergence/gap_crosswalk.py build --check
exit 0: crosswalk up to date

uv run --no-project --no-config --no-env-file --no-build --no-managed-python --no-python-downloads --offline --with typesafe-sdk==0.7.2 python3 -m unittest discover -s tests -p 'test_gap_crosswalk*.py' -v
exit 0: 22 tests, no skips

python3 scripts/component_matrix.py --write
exit 0; outputs unchanged
python3 scripts/new_host_grand_list.py --write
exit 0; outputs unchanged
python3 scripts/validate.py
exit 0 after the final P3 registry commit; final commit is registry-only
git diff --check
exit 0

The first eight-test stub run exited 1 because three preloaded mock responses lacked elapsed metadata. The final streamed fixture exercises native read/close timing; that failed attempt and correction remain in evidence/artifacts/typesafe-sdk-transport-20261005/checks.json, with final returned output alongside it. Independent review also corrected the missing-SDK hint to preserve the original command's required --set/--out arguments; no material findings remain.

Decision record

docs/decisions/2026-10-05-typesafe-sdk-transport.md records the three gates, retry comparison, header/timing changes, alternatives, overturn conditions and completeness critic.

Host evidence

Not applicable: no host receipts changed.

Checklist

• No GitHub Actions changes; existing offline CI build remains stdlib-only.
• No secrets printed or committed; no new required secret.
• No new paid hosting, subscription or billing surface.
• Peer-owned files and Withdrawn NativeStack2604 currency qualification handoff #721's worktree preserved.
• Registry committed last; required post-commit validation passed.