[nik13]

Merge hold: fixed instrumentor release required

Do not merge this route yet. Independent recall confirmed cancellation, explicit-close and incomplete-stream telemetry defects in the latest published traceAI-openai==0.1.12 / fi-instrumentation-otel==1.2.2. The newly checked-in lifecycle gates deliberately fail against those releases; no xfail, skip or continue-on-error hides the blocker.

Reuse #520 and #536. #536 at 93907f3cfbbd39f18a6e3a3ad11171757b81057e includes #520 at 3e19fa3b2333f812228b62fade8450a57a0fa798; its built packages pass every PrismML contract. This PR does not copy those shared fixes or create a second instrumentor.

Before unblocking: land/release the shared repairs through their own gates, confirm the real registry versions, update this recipe's requirements and README plus future-agi/docs#1097 and its workflow/static contracts, then rerun both source and published-package matrices. No fixed version has been invented or published here.

Summary

• Refresh the existing PrismML Bonsai recipe to OpenAI 3.42.1 and already-published tracing packages; retain OpenAI 1.69.0 compatibility coverage.
• Require model metadata on streamed and failed calls; execute the exact README setup with real loopback OTLP export.
• Fix the application, independently of tracing: close its stream via a context manager, require choice 0's finish reason, and exit 1 with an actionable message after EOF or premature [DONE] without completion. Partial printed text is not a successful answer.
• Add strict request/stream interruption, async cancellation, explicit sync/async close, incomplete stream, terminal-reason and usage-tail contracts. Real subprocess tests send SIGINT during the actual CLI's request and streaming phases.
• Keep native-first maintained-package setup, endpoint/key separation, privacy controls and concise customer instructions. No copied telemetry helper, second span or live-vendor claim.

Recall reproduction and validation

All tests/builds ran in the outbound-denying macOS sandbox, with isolated HOME, synthetic credentials and loopback/MockTransport model traffic.

Packages	Python / OpenAI	Recipe + exact guide
Published pins, initial recall contracts before application fix	3.10 / 1.69.0	10 failed, 106 passed
Published pins, initial recall contracts before application fix	3.13 / 3.42.1	10 failed, 106 passed
Published pins after application fix	Both bounds	11 failed, 120 passed
This PR's own built packages	Both bounds	11 failed, 120 passed
Built shared-repair #536 packages	Both bounds	131 passed

The four application-only incomplete-stream contracts changed from failure to success. A stream context also fixes the CLI's interrupted-stream export, but request cancellation, standalone close and premature-stream telemetry remain release-blocked. The 11 residual failures are retained as release gates.

Full #536 OpenAI package suite: 803 passed, 8 SDK-feature skips at the floor; 811 passed at latest. Core cancellation regressions: 68 passed at each bound. Optional Fireworks/vLLM compatibility suites were included. An initial floor package run exposed a test-transport mismatch from leftover optional httpx2 after downgrading OpenAI; removing that latest-only leftover restores the clean-floor results above. It was not an application defect or a suppressed test.

Current dev b059859f is an ancestor. Dependency checks pass; release-plan validation covers 129 packages; diff checks pass. No lockfile, package version or release configuration change. No registry publication, workflow dispatch or merge.

Related guide and E2E coverage

Companion guide: future-agi/docs#1097 — also held on the genuine fixed release. Its exact cancellation example is now a failing published-dependency gate rather than an unsupported success claim.

Not applicable to the product E2E harness: this is an SDK recipe. Real CLI entry points, HTTP/OTLP transport and exact customer snippets are exercised. No live Bonsai inference or Future AGI Cloud readback is claimed.

🤖 Generated with Claude Code