Remove TemporalIdGenerator.seed_span_id and seed_trace_id

[DABH]

Stacked on #1868, and CI here inherits its base's current uv sync failure (the two-week exclude-newer window for temporalio-openai-agents 1.0.0 clears on 2026-09-30) until that PR is updated or lands. Follow-up to #1856: its OpenAI Agents half landed as temporalio/ai-integrations#26, and this is the OTel half that was left behind when that PR closed.

What was changed

Removes TemporalIdGenerator.seed_span_id() and seed_trace_id() from temporalio.contrib.opentelemetry, together with the context-local seed queues behind them, the branches in generate_span_id() / generate_trace_id() that consumed them, and the one test that only exercised the seeding. ReplaySafeTracerProvider.id_generator() stays. Net −80 lines of code, no additions.

💥 Breaking change

Code that calls either method raises AttributeError after upgrading. TemporalIdGenerator is not exported from temporalio.contrib.opentelemetry, lives in a private module, and carries the experimental warning, but it is reachable through ReplaySafeTracerProvider.id_generator() and the provider's constructor, and the two methods were public-named and documented — so this is a removal, not a refactor, and it is logged under Breaking Changes next to #1868's entry.

Who is affected: the API was added in #1286 (1.24.0) for exactly one caller, the OpenAI Agents OTel interceptor, which #1868 removes from this repo. The standalone temporalio-openai-agents does not use it. A GitHub code search finds one consumer outside vendored SDK copies: temporal-community/temporal-agent-harness, which carries a copy of the old interceptor.

Why remove it rather than keep it

• It was the wrong mechanism, and the fix that makes it unnecessary has shipped. Seeding made the worker mint a span under the caller's IDs — a second copy of the caller's span, exported from a second process — which is what produced the duplicated roots and missing parents in openai_agents OpenTelemetry tracing: detached spans and missing client root span when client and worker share a process #1852 / openai_agents OTel bridge logs "Failed to detach context" whenever a temporal:startActivity span finishes #1853. Keep OpenAI Agents traces connected across Temporal execution #1856 and ai-integrations#26 replaced that with the propagated IDs as a remote parent (SpanContext(..., is_remote=True)), the standard OpenTelemetry way, which needs no ID seeding. Keeping the API keeps the anti-pattern one import away.
• It is a footgun on its own. A seed sits in a context-local queue until the next generate_* call in that context, whatever span that turns out to be, so a seed that is not consumed where it was intended yields two spans with the same ID.
• After Remove OpenAI Agents integration #1868 there are no callers and no coverage. The one remaining test checked the queue's context-locality, not any use of it.
• Why now. The module is experimental and documented as such; Remove OpenAI Agents integration #1868 already carries a Breaking Changes entry for the OpenAI Agents move in this release, so users get one migration note instead of two a release apart; and the alternative is shipping a release where the methods exist, are unused, and are untested. Tim asked for exactly this removal in #1856 review.

Migration: connect a trace across a boundary by setting a remote-parent SpanContext, as temporalio-openai-agents does. To control the ID sequence itself, pass your own IdGenerator to create_tracer_provider(id_generator=...); TemporalIdGenerator wraps it and defers to it outside workflows.

Checklist

1. tests/contrib/opentelemetry: 75 passed on this branch. test_temporal_id_generator_seeds_are_context_local removed (it only tested the seeding).
2. Lint gate clean on the changed files.
3. CHANGELOG entry under Breaking Changes.

[brianstrauch]

+1, but would recommend shortening the changelog

[DABH]

@brianstrauch done, thanks