experiment: prove stock Code Mode observation boundary

[bateau84]

Parent

Child of #45. This PR is intentionally limited to the remaining Code Mode inner-call observation question.

Result

Stock OpenCode 2.0.23 supports a useful partial same-process observation path without any OpenCode patch:

• allocate a unique per-inner invocation ID in a supported tool.transform leaf wrapper;
• capture the actual effective tool;
• capture decoded/executable input;
• bind the inner call to the real outer execute Session/message/CallID carried by Tool.Context;
• preserve start and handler-terminal ordering across identical concurrent calls and reverse completion.

It does not expose enough supported public surface to capture the exact terminal value/error seen by the Code Mode script:

• all public inner tool hooks reuse the outer execute CallID, so concurrent identical calls cannot be uniquely correlated there;
• the transformed handler returns before later core hooks/normalization and Code Mode output conversion;
• the post-conversion success value is visible only to the private @opencode/codemode tool.after hook wired to OpenCode's internal progress hook;
• a failed tool promise is materialized into the JavaScript error bound by catch later in the interpreter, with no plugin hook at that boundary.

The final caller-terminal capability is therefore reported as unsupported. No earlier value is promoted or reconstructed.

Provider-free stock-runtime proof

The actual-image integration probe uses a loopback deterministic OpenAI-compatible fixture provider and stock 2.0.23. It covers:

1. one successful inner call;
2. an inner thrown error caught by Code Mode;
3. two identical concurrent calls;
4. reverse completion order;
5. multiple calls under one outer execute;
6. collector-shaped outer script output that must not fabricate an inner observation.

It also mutates one result after the transformed handler returns. The Code Mode script receives the mutated value, proving the handler terminal is not caller-final.

The probe remains diagnostic only and reports:

{
  "status": "unsupported",
  "evidence_eligible": false
}

PR #41 reuse

Only the useful test shapes and counterexamples were retained. None of PR #41's patched OpenCode observer seam is reused.

Scope

• stock OpenCode 2.0.23 only;
• Code Mode only;
• no fork/patch;
• no isolation/signing architecture;
• no changes to Loom scoring;
• no claim that target-writable probe diagnostics are authoritative evidence.