chore(ci): run compliance harness 1.13.1 once per codec

[eli-r-ph]

💡 Motivation and Context

• The compliance harness skips the AI lane and event options for this SDK. The adapter does not advertise either capability, and the harness version predates both suites.
• The adapter also folds options back into legacy $ properties. Since the options PRs earlier in this stack, the SDK takes options= directly, so this fold tests a path users no longer need.
• The reusable harness workflow posts a PR comment, which needs pull-requests: write. That step failed on #1019 for reasons unrelated to the SDK.

This PR runs harness 1.13.1 with every capture suite the SDK supports, once per codec.

Changes

• The adapter advertises capture_ai_v1 and event_options. Both suites now run against this SDK.
• /capture and /capture_ai share one handler. It passes options, properties and a supplied uuid to the SDK unchanged, and returns the uuid the SDK sent.
• The adapter reads its codec from COMPRESSION (gzip, deflate or zstd, default gzip). It advertises only that encoding_<codec> capability and uses the codec on both lanes. An unknown value stops the adapter at startup.
• The harness has one enable_compression flag but one test per codec, so CI runs one job per codec. There is no brotli job, because the SDK has no brotli support.
• The workflow runs the harness and adapter containers directly, in the same shape as posthog-go. The reusable workflow cannot pass COMPRESSION to the adapter.
• The report goes to the job summary and an artifact, not a PR comment. The workflow now needs only contents: read.
• The adapter image and the adapter unit-test job install the zstd extra. docker-compose.yml uses harness 1.13.1 and passes COMPRESSION.

No SDK code changes. A behavior change for the harness only: /capture now forwards a supplied uuid, as /capture_ai already did.

💚 How did you test it?

• The harness 1.13.1 ran locally once per codec. Every suite ran (capture v1, AI capture v1, flags, local evaluation), and every test passed: 223 for gzip, 221 for deflate and zstd. The gzip run has two more tests because gzip has more encoding tests.
• New adapter tests:
  • test_health_and_init_use_one_codec catches an adapter that advertises the wrong encoding_ capability, or that sets the codec on only one lane.
  • test_capture_passes_options_and_uuid_unchanged catches an adapter that rewrites options, drops the uuid, or returns a uuid other than the SDK's.
• Break-on-purpose: each of these broke the new tests: a hard-coded encoding_gzip, options dropped, the AI lane codec left unset.
• Not checked: the CI matrix itself. It runs on this PR.

📝 Checklist

• I reviewed the submitted code.
• I added tests to verify the changes.
• I updated the docs if needed.
• No breaking change or entry added to the changelog.

If releasing new changes

• Ran sampo add to generate a changeset file

None needed: this changes CI and the compliance adapter only.

🤖 Agent context

Autonomy: Human-driven (agent-assisted)

Written with Cursor (Claude Opus) under the direction of the assignee. Skills used: writing-pr-descriptions.

Agent calls worth review:

• One adapter instance per codec, chosen by an env var, with the matrix in the workflow. The alternative, one job that restarts the adapter per codec, keeps one report but makes failures harder to read.
• The default codec is gzip, so docker compose up without COMPRESSION behaves as before.

[eli-r-ph]

@greptileai review

[dustinbyrne]

The expanded compliance matrix and adapter forwarding look sound. Existing CI exercises both capture lanes across all three codecs. One small documentation follow-up: the standalone adapter-test setup needs the zstd extra.

AI-assisted review.