feat: add conversation materialization

[Amidwestnoob]

PR draft 2: conversation materialization

Tracks VectifyAI#224

Status

This change is stacked on VectifyAI#223. It is independent of VectifyAI#179.

Summary

Add a small transport-neutral conversation package.

The package converts caller-supplied closed conversation windows into immutable source documents. It keeps source, speaker, authority, time, and evidence identifiers. It writes content hashes, closed-window fingerprints, and a replay-safe manifest.

The caller remains the source of truth. The wiki remains a read-only projection.

Changes

• Add typed conversation turn and window records.
• Validate ISO timestamps with UTC offsets.
• Sort turn and compile times by normalized UTC values.
• Preserve caller order when normalized turn times are equal.
• Normalize text with caller-supplied pure hooks.
• Fail closed for oversize turns.
• Apply the document limit to final rendered characters.
• Provide explicit middle truncation with a digest.
• Deduplicate by (conversation_id, turn_id).
• Reject conflicting duplicate identities.
• Split only at turn boundaries.
• Verify exact ordered turn coverage after every split.
• Stop before mutation when one pending recovery turn cannot fit.
• Render deterministic source-document bytes.
• Store a canonical fingerprint for every closed window, including an empty window.
• Store a fingerprint for each active document set.
• Reject a changed normal replay before it writes a new path.
• Use collision-resistant path tokens.
• Validate manifest paths, UTF-8 text, timestamps, hashes, windows, and completed hashes.
• Verify manifest file hashes.
• Select oldest pending compile inputs without starting compilation.
• Support custom manifest names during compile selection.
• Store optional compile receipts by content hash.
• Validate bounded compile budgets.
• Re-split only uncompiled documents with a smaller recovery limit.
• Match recovery turns with the full stable identity.
• Remove replaced uncompiled files only after the new files and manifest are durable.
• Use the shared mutation lock for manifest updates and verified claim updates.
• Apply caller-verified claims through the feat: add temporal claims to wiki pages VectifyAI/OpenKB#223 claims API.

Scope boundaries

This change does not depend on VectifyAI#179. It does not copy the ingest-bundle pipeline.

This change does not add source-specific rules, live input, source data, scheduler wiring, or model calls.

A future VectifyAI#179 adapter can call this materializer.

Tests

The synthetic tests cover cleaning, hook order, explicit truncation, rendered limits, first-turn fit failure, exact split coverage, duplicate handling, UTC ordering, equal-time caller order, empty-window identity, changed segment sets, collision-resistant paths, manifest validation, corrupt UTF-8 text, custom manifest selection, compile receipts, concurrent manifest updates, concurrent claim updates, recovery identity, recovery no-mutation failure, stale-file removal, budget checks, verified authority, and claim idempotence.

Verification receipt

• Earlier blocking review reproduction: PASS (7 grouped checks reproduced 7 failures)
• Final review reproduction: PASS (5 checks reproduced 5 failures)
• Targeted regression tests: PASS (34 passed)
• Full python -m pytest: PASS (1,320 passed; 2 warnings in existing removal tests)
• python -m ruff check .: PASS
• python -m ruff format --check .: PASS (130 files already formatted)
• python -m mypy on all new modules: PASS
• git diff --check: PASS
• tests/test_file_size.py: PASS (5 passed)
• Untracked-file diff check: PASS
• Forbidden-string scan: PASS
• feat: add temporal claims to wiki pages VectifyAI/OpenKB#223 file scope scan: PASS
• uv build: PASS (source archive and wheel)

Review notes

The package uses shared OpenKB lock, atomic write, frontmatter, and claims helpers. It does not modify openkb/claims.py or openkb/agent/compiler.py from VectifyAI#223.

The deployment values in the public document are examples. They are not code defaults.

Residual risk: A process stop after a recovery file write and before the manifest write can leave an untracked recovery file. It is not selectable. Automatic cleanup is outside this change.