Complete 0083 with the token-budget failure path

[chris-colinsky]

Third and final PR for proposal 0083 (per-prompt token-budget observability). Adds the structured-output-failure path and marks 0083 implemented. Follows PR A (completion path) and PR B (WARNING surfaces).

The failure path needed no source change

The budget already evaluates on any usage-bearing per-attempt event, so the failure path falls out of the generic handling from PR A/B:

• A structured_output_invalid failure carries usage (proposal 0082), so it drives the OTel span attribute, the two §11 metrics, and the §7 WARNING log exactly like a completion, and the Langfuse ERROR Generation carries metadata.token_budget.
• A no-usage failure (provider_unavailable) populates token_budget on the event but records nothing token-budget (there is nothing to compare against).

PR C is therefore harness wiring plus the conformance flip:

• test_observability.py — route fixture 131 to the token-budget runner; add expected_error handling (pytest.raises + raised_from, with the §7 category riding the span status_description); make the exact-metrics assertion conditional on a declared metrics: block (a failure still records the baseline operation.duration); implement the two without-usage invariants; and fail loudly on an unmapped invariant name.
• test_fixture_parsing.py — un-defer 131.
• conformance.toml — 0083 → implemented (all six fixtures 126-131 run).

Whole-proposal review

Ran a final multi-agent adversarial review over the complete A+B+C impl (0 blockers). Folded in: getattr parity in the Langfuse metadata block, the invariant-name guard, and end-to-end coverage the per-PR reviews could not reach — a structured-over-budget failure through Langfuse, the PromptManager.render → PromptResult.token_budget propagation, the langfuse extra-key tolerate case, and the filesystem unified malformed-budget path. Two non-blocking spec questions (Langfuse failure-path exceedance discoverability; the extra-key policy divergence) are with spec.

Tests

Fixture 131 (both cases) plus the folded-in unit coverage. Full suite green; ruff and pyright clean.

[Copilot]

Pull request overview

Completes proposal 0083 (per-prompt token-budget observability) by wiring and asserting the failure-path behavior in the conformance harness, and flipping the conformance registry status to implemented. This rounds out token-budget signals across both successful completions and usage-bearing failures, while explicitly asserting “no usage, no evaluation” semantics.

Changes:

• Extend the conformance observability runner to execute fixture 131 (token-budget on structured-output failure), including expected_error handling and new “without usage” invariants.
• Add unit coverage for PromptManager.render token-budget propagation/copy semantics, unified-backend malformed-budget fallback eligibility, and Langfuse failure-path WARNING-vs-ERROR precedence.
• Mark proposal 0083 as implemented in conformance.toml and remove the 131 parsing deferral.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
tests/unit/test_prompts.py	Adds unit tests for token-budget propagation and unified malformed-budget fallback behavior.
tests/unit/test_observability_langfuse.py	Adds Langfuse test asserting structured-output over-budget failures stay ERROR (no WARNING override) while preserving token_budget metadata.
tests/conformance/test_observability.py	Routes fixture 131 through the token-budget runner; adds expected_error handling, conditional metrics assertions, and invariant-name guarding.
tests/conformance/test_fixture_parsing.py	Removes the deferral for fixture 131 now that the harness supports it.
src/openarmature/observability/langfuse/observer.py	Makes token_budget metadata extraction defensive via getattr, matching existing OTel/helper behavior.
conformance.toml	Flips proposal 0083 to implemented and documents the completed A/B/C rollout.