chore(release): 0.203.2

[github-actions]

Release-time preparation from merged main. Feature PRs do not carry version or changelog edits. A maintainer must approve the bot-created Actions workflow run before its checks can run.

[tangletools]

✅ Auto-approved PR — 8f61d65f

Blanket team auto-approval is intentional. The merge gates are CI and codex-p1.
No automated review runs on this PR. This approval rests on the rule above alone.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-10-01T17:38:39Z}

[drewstone]

Independent review: exactly five generated release files; four version fields move0.203.1→0.203.2 and changelog names the merged attempt-retention fix. Production source, dependencies and workflow gates are unchanged. Source908 and assertion repair910 are merged; actual public203.1 baseline loses both retry/resume bytes while the source consumer preserves them. Merge releases metadata; publication and exact public-artifact consumer proof remain separate and owned by the existing release lane.

[drewstone]

Public 0.203.2 delivery is complete.

• Immutable source/tag: 8517510af50604e8ab64b28d8a13acd38215aafe / v0.203.2.
• Maintained publisher 36904773087 completed successfully for verification, npm and PyPI.
• Public npm tar: 4,759,539 bytes, SHA256 a0064302ec6d47bad3f800adf2708249243d97b3639058bffcf21f33e38f4bbd. Native signature audit passes. Attestation subject, source, tag, workflow and run match.
• All 716 installed regular files match the archive by SHA256, length and bytes. Archive/installed permissions remain separately recorded. Five executable source-map contents match exact release Git.
• Existing operational consumer passed through both a fresh public install and Lab's actual install: retry/resume preserve both attempts, cache reuse performs no new callback, four legacy files stay byte-identical, outside-attempt writes are refused, and optimization resume recovers its failed candidate.
• Both PyPI artifacts pass cryptographic attestation verification. All 16 Python source files match release Git. The fresh official-index Client installs/imports at 0.203.2.

The actual public 0.203.1 baseline lost the first artifact and trace in both retry and resume. Those records remain preserved.
The real npm propagation 404, default Python-index refusal and corrected source-path/interpreter observations also remain preserved.

Full receipts and native SDK records: /mnt/traces/eval-release-02032-20261001 on GTR.
Frozen 257-file evidence manifest, plus manifest itself: /mnt/traces/eval-release-02032-20261001.receipts.tgz, 5,691,398 bytes, SHA256 566492f9f4e395282214a01bccc8d662e5ad147f05158035a7d8b4322952dcd8.
Installed dependency/virtual-environment directories are excluded; exact public artifacts, package locks, file identities, native records and import receipts are retained.

Lab adoption PR1220 pins the qualified version without an intermediate 0.203.1 pin.
This is SDK filesystem qualification: zero provider/model requests, zero Sandbox allocations, zero research units.
It does not establish all-harness capture, billing completeness, subscription readiness or scientific outcomes.
No new unit tests or local suites were added/run; the maintained publisher's existing checks ran.