feat(import): conversation-export importers — reconstruct proposals from prior agent history

[plind-junior]

vouch has register_source / register_source_from_path for citing a source, an Obsidian vault sync for editing approved pages, and a proposed bulk-markdown importer (#321). what it lacks is a path for someone arriving with existing agent history — chat exports, prior-tool memory dumps — to seed a kb from it. today that user starts empty and re-teaches everything by hand, which is the biggest friction on adoption.

this proposes format-specific importers that parse an export dump and emit proposals (claims/pages) into the review queue — same gate, nothing auto-approved.

proposed surface

vouch import <format> <path> [--dry-run] [--max-proposals N]
# formats: chat-json | markdown-vault | memory-export

• per-format readers that normalize an export into candidate claims/pages, then route each through proposals.propose_*. reuse the extraction path feat: bulk corpus import — propose pages for a directory of markdown #321 builds for directory-markdown; add adapters for json chat exports and generic memory dumps.
• --dry-run reports would-propose counts (reuses the existing dry_run= flag on propose_claim/propose_page) without touching the queue.
• --max-proposals caps a single import so a large history can't flood review in one shot.
• de-dupe candidates against existing approved artifacts via find_similar_on_propose (feat: propose-time similarity warnings for duplicate claims #147) at propose time.

review gate & scope

every imported item is a proposal — the importer has no path to approve. a human drains the queue exactly as for any other write; the gate is untouched. --max-proposals and dedup keep an unattended import from overwhelming the reviewer. local-first: reads a local file, writes only pending proposals into .vouch/, no network, no hosted component.

acceptance criteria

• vouch import <format> <path> emits pending proposals via propose_*; no auto-approve.
• --dry-run reports counts without enqueuing.
• --max-proposals N caps proposals per run and reports when the cap is hit.
• imported candidates are de-duped against approved artifacts via feat: propose-time similarity warnings for duplicate claims #147.
• at least two formats (chat-json + markdown-vault) land, sharing feat: bulk corpus import — propose pages for a directory of markdown #321's extraction core.
• tests/test_import.py covers the propose-only invariant, the cap, and dedup.
• make check green.

related: #321 (bulk markdown import — this generalizes it to export formats), #147 (dup warnings), #319 / #320 (other propose-only ingest paths).