MCP-native harvest: post a transcript instead of rsyncing session files

[raphasouthall]

Part of #115.

What

vault_harvest_transcript(transcript, session_id, source_agent, dry_run) — the client posts its own session text and the server harvests it, so capture no longer depends on the server being able to read the client's filesystem. Adds OmpProvider for ~/.omp/agent/sessions/*/*.jsonl so the file-scan path and backfills understand omp sessions too.

How

The classify/redact/dedup/save loop moved verbatim out of harvest_sessions into _harvest_messages, which both entry points call. harvest_sessions keeps session discovery, the mtime state filter and the report; its public behaviour and report shape are unchanged and no existing test was touched. Because redact_secrets sits inside the shared seam, ahead of the dedup check, the #113 contract applies to the new unattended write path by construction rather than by a second call site.

harvest_transcript writes the posted text to a temp file and hands it to the named provider's existing extract_messages, so no provider grew a second entry point. The temp file is unlinked in a finally covering both the write and the parse.

No server-side chunk assembly. A transcript over MAX_TRANSCRIPT_BYTES (4 MiB) returns an error telling the client to split on newline boundaries: JSONL chunks are whole messages, each chunk harvests independently, and the cosine dedup absorbs overlap. The re-post guard reuses harvest_state under an mcp:<source_agent>:<session_id> key holding the transcript's sha256, so a retried POST is a no-op; the mtime comparison in harvest_sessions is now numeric-only so the two key namespaces cannot be confused.

OmpProvider keeps only user and assistant roles — toolResult also carries text parts, and including them drowns the pre-filter in tool output — and only text content parts, dropping thinking and toolCall.

Gate

uv run ruff check src/ tests/ exit 0. uv run pytest -q exit 0, 804 passed (792 existing + 12 new), reproduced cold after the build.

Live verify

Against real on-disk omp sessions, read-only: find_sessions(3) returned 3, and the newest parsed to 86 messages (31 user, 55 assistant) with all 166 toolResult messages and every thinking/toolCall part excluded; a dry-run harvest of that file produced 8 insights. Through the tool registry, a dry-run post of a Claude-shaped line returned counts {'bug': 1}, and an unknown source_agent returned an error naming the six registered providers. End-to-end verification against the deployed server from both a Claude Code and an omp client follows the merge, and the interim rsync timer is retired only after that.

Known gap

The omp tree also holds depth-3 <project>/<session>/<AgentName>.jsonl subagent transcripts. The glob is deliberately depth 2, so those are not picked up by the file-scan path; they are largely tool chatter and would multiply the scan. A client can still post one explicitly.