Add 0083 token-budget WARNING surfaces

[chris-colinsky]

Second of three PRs for proposal 0083 (per-prompt token-budget observability). Adds the two vendor-neutral over-budget WARNING surfaces on top of PR A's completion-path signal. The structured-output-failure path (fixture 131) and the implemented flip follow in PR C.

What it adds

• Langfuse (§8.4.3): on a budget exceedance the successful Generation ends at level = "WARNING" with a statusMessage naming each breached bound (e.g. token budget exceeded: input 20 > 10), and metadata.token_budget carries the declared bounds. A hard ERROR still takes precedence on the failure path (the advisory WARNING only applies to a successful call).
• §7 log: one WARNING record per over-budget attempt on the openarmature.observability logger, naming the prompt and the breached bound(s). It is its own surface, dispatched independently of the span (disable_llm_spans) and metric (enable_metrics) gates. The record correlates via correlation_id; it does not carry OTel trace_id/span_id, since OA's observer never activates its spans and the record is emitted outside any span context.
• Shared helper: the per-bound evaluation used by the span, metrics, Langfuse WARNING, and §7 log is lifted to observability/llm_event.py so both observers share one implementation.

Conformance

Un-defers fixture 130 (the Langfuse WARNING). The §7 log is prose-pinned (no fixture) and covered by unit tests. conformance.toml stays not-yet until PR C lands the failure path.

Tests

Fixture 130 plus unit coverage for: the Langfuse WARNING (over/under budget, both-bounds statusMessage, metadata.token_budget on success and on a failed Generation) and the §7 log (prompt identity, multi-bound formatting, per-attempt cardinality, the structured-output-failure and no-usage cases, and independence from both the span and metric gates). Full suite green; ruff and pyright clean.

[Copilot]

Pull request overview

This PR extends OpenArmature’s proposal 0083 token-budget observability by adding two vendor-neutral “over budget” WARNING surfaces (Langfuse Generation level/statusMessage and a spec §7 warning log), and centralizes shared per-bound evaluation logic so observers cannot drift.

Changes:

• Add §7 token-budget WARNING logging on openarmature.observability, emitted once per over-budget attempt and independent of span/metric gates.
• Add Langfuse advisory WARNING mapping for successful completions that exceed declared token budget bounds, including statusMessage formatting and metadata.token_budget.
• Move the shared token-budget per-bound evaluation helper into observability/llm_event.py and enable conformance fixture 130.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
tests/unit/test_observability_otel.py	Adds unit coverage for the §7 warning log surface (cardinality, formatting, gating independence, prompt identity, failure/no-usage cases).
tests/unit/test_observability_langfuse.py	Adds unit coverage for Langfuse WARNING level/statusMessage and token budget metadata behavior on success and failure.
tests/conformance/test_observability_langfuse.py	Wires token_budget fixture data into the conformance harness prompt objects and enables fixture 130.
tests/conformance/test_fixture_parsing.py	Removes the “not implemented” deferral for fixture 130 while keeping 131 deferred.
src/openarmature/observability/otel/observer.py	Implements the §7 warning log surface and switches to the shared token-budget evaluation helper.
src/openarmature/observability/llm_event.py	Introduces the shared _token_budget_evaluations helper for consistent per-bound evaluation across observers.
src/openarmature/observability/langfuse/observer.py	Implements Langfuse Generation WARNING level/statusMessage on budget breach and maps metadata.token_budget.
conformance.toml	Updates proposal 0083 notes to reflect fixture 130 now running while 131 remains deferred.