test(http-recorder): match cassettes prompt-agnostically for transport tests

[jack-champagne]

The three session.llm native recorded scenarios are the only failures on local/amicode — the base branch's test job reports 5 pass / 2 fail and those three are the entire cause. So the fork's base is red, every PR into it inherits a red check, and nothing in the fork can be validated. #93 is the current casualty: its own suite is 3366 pass / 3 fail, with zero failures in its own files.

Cause is ours

AMICODE-PATCHES.md ADDENDUM 2 consolidated the prompt so system.ts provider() returns [providerBase(model), PROMPT_COMMUNICATING], appending communicating.txt for every model family. The cassettes still record the pre-patch prompt, so all three break — on a diff of prose.

Why not just re-record or patch the fixtures

Re-recording needs live credentials for all three providers (OpenAI API key and an OAuth JSON blob, Zen console token + org id, Anthropic key) and makes real API calls.

Patching the recorded prompt text would work until the next prompt edit — and that patch already carries three addenda. It would be the fourth fix to the same recurring break.

What this does instead

Those scenarios assert transport mechanics: that a tool loop is driven to a final text answer. The prompt is incidental, so matching on it couples them to text they don't test.

RecorderOptions.match is an existing seam (types.ts:85, threaded through http() into recordingLayer) that nothing used. This adds promptAgnosticMatcher alongside defaultMatcher and passes it from this one test.

The prompt arrives in three different shapes, all covered:

shape	provider
instructions	OpenAI Responses
system[]	Anthropic Messages
input[].role == "system"	the OpenCode proxy

System-message content is replaced with a placeholder rather than deleted, so "a system message is present, with content" remains part of the match.

Verified narrow

The risk with a laxer matcher is that it stops catching real differences, so test/matching.test.ts pins both directions. It ignores prompt prose in all three shapes, and still rejects:

• a different user message
• a different model
• a different tool set
• a different URL
• a missing-vs-present system message

Results: http-recorder 49 pass / 0 fail; the three recorded scenarios go 3 fail → 3 pass; typecheck green across 23 packages.

Scope

Test infrastructure only — no src/ file outside http-recorder imports this package, so nothing shipped changes. Deliberately does not touch requestDiff, so failure messages are unchanged.