feat: record compiler version provenance

[MaximGit-tech]

Summary

Adds compiler provenance metadata to every newly compiled memory so Statewave can tell which compiler configuration produced it.

This implements the missing compiler-version information requested in #443.

What changed

LLM compiler

LLM-produced memories now record:

• compiler
• model
• prompt_version
• compiler_version

prompt_version is a deterministic SHA-256 fingerprint of the prompt templates used to produce the memory.

The normal extraction pass versions the active _system_prompt(), including the compose-unit variant when enabled.

Recall-sweep memories are versioned separately. Their prompt version includes both the active main extraction prompt and _RECALL_SWEEP_PROMPT, because the sweep output also depends on the first-pass results it receives.

Recall-sweep memories continue to carry:

"pass": "recall_sweep"

Heuristic compiler

Heuristic-produced memories now record:

{
    "compiler": "heuristic",
    "compiler_version": "heuristic:1",
}

The heuristic compiler uses an explicit revision rather than hashing Python source code, so formatting or comment-only changes do not unnecessarily change the compiler version.

Existing claim metadata is preserved and merged with the new compiler provenance.

Structured candidates

Memories produced directly from structured memory candidates now record:

{
    "compiler": "structured",
    "compiler_version": "structured:1",
}

This keeps their provenance accurate because structured candidates bypass both LLM extraction and heuristic extraction.

Producer-supplied metadata is preserved, while the reserved compiler provenance fields are set by Statewave.

Shared versioning

Added a small compiler-version helper that produces deterministic SHA-256 fingerprints from version components.

Components are encoded separately before hashing so different component boundaries cannot accidentally produce the same fingerprint.

Backward compatibility

No database migration or backfill is required.

Existing memories without compiler_version remain unchanged and therefore continue to represent unknown compiler provenance, as described in #443.

Tests

Added regression coverage for:

• LLM compiler provenance
• compiler-version changes when the model changes
• compiler-version changes when the active prompt changes
• recall-sweep provenance
• heuristic compiler provenance
• preservation of heuristic claim metadata
• structured candidate provenance
• protection of reserved structured compiler metadata

Also updated existing heuristic claim tests to account for compiler provenance being present even when no claim is attached.

Validation:

1169 passed, 40 skipped, 3 warnings, 8 subtests passed
ruff check . -> All checks passed

The remaining warnings are pre-existing AsyncMock warnings unrelated to this change.

Closes #443

[smaramwbc]

Nice work, this is thorough. All three write paths are covered, the tests bite, and no migration is needed since metadata_ is already JSONB. Merging, and I am following up with one tweak: the structured path reserves compiler and compiler_version but not model or prompt_version, so a candidate could still supply those two. Thanks!