feat(doctor): --classify-orphans pass for un-typed low-prior beliefs (#206)

[robotrocketscience]

Closes #206.

Summary

Adds `aelf doctor --classify-orphans` — a targeted reclassification pass that finds beliefs whose `type` was never resolved AND that have never received any feedback, then re-classifies them through the same Haiku batch path `aelf onboard --llm-classify` already uses. Strict superset of onboard: only adds calls, doesn't replace anything.

Orphan definition (both signals required)

• `type = 'unknown' OR type IS NULL` — never successfully typed by onboard/ingest.
• `alpha + beta <= 2` — untouched prior is alpha=1, beta=1 (sum=2); any feedback event pushes the sum above 2.

The other orphan signals from the issue (zero non-CONTAINS edges, entity_index miss) are noted in code comments as future extensions once #143 / edge-type auditing land.

CLI

• `aelf doctor --classify-orphans` — full pass.
• `--dry-run` — count orphans + show before-distribution; no LLM calls, no DB writes.
• `--max N` — cap classifications per run (recommended: 500 for large stores).
• Reports orphans found, classified, skipped, before/after type distribution, input/output token counts, request count, fallback count, and an estimated cost line.

Implementation notes

• Reuses `aelfrice.llm_classifier.classify_batch` — no new LLM client.
• Updates only `type` and `origin` (set to `agent_inferred`); `alpha`, `beta`, lock state, `demotion_pressure`, and timestamps are preserved.
• Auth failures and token-cap-exceeded errors propagate out and exit 1.
• When the whole batch falls back to regex, all orphans are marked skipped (regex-fallback is designed for fresh inserts in scan_repo, not in-place updates).

Test plan

• 14 new tests in `tests/test_doctor_classify_orphans.py`, all green; full suite 1598 passed / 5 skipped.
• `test_classify_orphans_95_percent_recovery` — synthetic orphan corpus, asserts >=95% recovery to non-unknown type.
• `test_classify_orphans_max_n_caps_llm_input` — verifies the SQL LIMIT actually bounds the LLM call.
• `test_cli_classify_orphans_no_api_key` — exits 1 with a useful message when ANTHROPIC_API_KEY is unset.
• `test_classify_orphans_dry_run_no_writes` — no DB writes, no network in dry-run mode.

Out of scope

• LLM-free orphan classification (separate spike).
• Auto-running on ingest — manual `aelf doctor` invocation only.
• `feedback_loop_synthesis` integration — defers to [v2.0] Write log as source of truth — append-only ingest_log + replay-capable derivation #205 landing.