Matrix intake: fuzzy investor matching + conversational in-thread edits (v0.1.0:86)

Close the two locked post-deploy enhancements for the Matrix intake bot.

Fuzzy matching (server-side, ships in the s9pk): new find_intake_candidates in
server.py returns ranked deterministic near-matches (difflib name similarity +
token-set Jaccard, legal-suffix-aware, + email Levenshtein <= 2); GET
/api/intake/match now returns {match, candidates}. The bot surfaces a numbered
shortlist so a near-duplicate (Charlie/Charles, Acme Capital vs Acme Capital LLC,
a one-char email typo) is confirmed by a human instead of silently creating a
second investor. Exact match still auto-attaches; fuzzy candidates are never
auto-attached. The optional LLM-judge re-rank is deferred.

Conversational edits (bot-side, ships on the Spark): any in-thread reply that
isn't yes/no/edit field=value is treated as a natural-language revision and
re-run through local Qwen (parse.revise). Email integrity is preserved -- a
changed address must literally appear in the instruction; the model's email
field is structurally unreachable. No-op revisions re-prompt.

Docs/current-state brought current; 27/27 backend tests green.

ROADMAP.md

@@ -100,9 +100,11 @@ Use the **matrix-bridge** repo's pattern to listen on a dedicated ten31-database
 - **CRM-side:** `POST /api/intake/investor` (service-auth) creates a new investor+contact **through the existing grid-save path** (so relational sync + audit + backup-on-write happen as with a UI edit; bot never does whole-blob RMW) or appends a meeting note to the interaction log for an existing investor; `GET /api/intake/match?q=` fuzzy-matches via the existing entity-resolution/email-matcher. New investor needs no fund at intake.
 - **Phases:** M1 = scaffold + parse + in-thread propose, **no writes** (proves Matrix↔Spark). M2 = intake endpoint + match + write-on-approve + tests. M3 (deferred) = business-card photo.
 
-**Post-deploy enhancement — fuzzy match + in-thread confirm (Grant, 2026-06-17).** Today `find_intake_match` is **exact-after-normalization** (`_normalize_text` = lowercase+strip), so near-misses — "Charlie" vs "Charles" (same last name), "Acme Capital" vs "Acme Capital LLC", a one-character email typo — return no match and the bot proposes a **new** investor, risking a duplicate the human approves without realizing a near-match exists. The existing in-thread approval gate is useless against this because the human is never *shown* the near-match. Fix: matcher returns **ranked fuzzy candidates** (deterministic pre-filter: normalized name similarity / token overlap + email edit-distance ≤ ~2), surfaced in-thread for the human to confirm or pick, with the **local Spark LLM optionally re-ranking/judging the shortlist** (good at Charlie/Charles + legal-suffix equivalence; fed only the shortlist, never the whole LP list). Keeps the approval gate but makes it effective against duplicates. Land **after** the live smoke — net-new logic + reply grammar + tests; the current exact match is safe and its failure mode (a duplicate) is recoverable via the existing entity-merge subsystem (`backend/entity_*.py`).
+**Post-deploy enhancement — fuzzy match + in-thread confirm (Grant, 2026-06-17). BUILT 2026-06-17 (v0.1.0:86), not yet deployed / live-smoked.** Today `find_intake_match` is **exact-after-normalization** (`_normalize_text` = lowercase+strip), so near-misses — "Charlie" vs "Charles" (same last name), "Acme Capital" vs "Acme Capital LLC", a one-character email typo — return no match and the bot proposes a **new** investor, risking a duplicate the human approves without realizing a near-match exists. The existing in-thread approval gate is useless against this because the human is never *shown* the near-match. Fix: matcher returns **ranked fuzzy candidates** (deterministic pre-filter: normalized name similarity / token overlap + email edit-distance ≤ ~2), surfaced in-thread for the human to confirm or pick, with the **local Spark LLM optionally re-ranking/judging the shortlist** (good at Charlie/Charles + legal-suffix equivalence; fed only the shortlist, never the whole LP list). Keeps the approval gate but makes it effective against duplicates. Land **after** the live smoke — net-new logic + reply grammar + tests; the current exact match is safe and its failure mode (a duplicate) is recoverable via the existing entity-merge subsystem (`backend/entity_*.py`).
+- **As built:** `find_intake_candidates` in `server.py` (deterministic — stdlib `difflib` name similarity + token-set Jaccard, legal-suffix-aware via `_strip_legal_suffix`, + email Levenshtein ≤ 2; ranked, ≥0.62, top 5). `GET /api/intake/match` now returns `{match, candidates}`. Bot: a new `_stage="disambiguate"` shortlist (`proposals.render_disambiguation` / `interpret_disambiguation` / `attach_to_candidate` / `promote_to_new`) — human picks a number / `new` / `no`. **The optional LLM-judge re-rank was deliberately deferred** (the deterministic filter already surfaces the named cases; an LLM judge is the right *pruner* for shortlist noise — build if the deterministic ranking proves too noisy in practice). Tests: `test_intake_endpoints.py` (server fuzzy cases), `matrix_intake/test_proposals.py` (disambiguation grammar), `matrix_intake/test_crm_client.py` (candidate shape).
 
-**Post-deploy enhancement — conversational (LLM-mediated) edits (Grant, 2026-06-17).** Today an in-thread correction uses a rigid grammar (`edit field=value`). Let a free-form reply that isn't `yes`/`no`/a literal `edit …` be treated as a natural-language revision instruction: send {current proposal + the instruction} back through local Qwen (`spark.py`, the same parse leg — no Claude, no scrub) and re-render the revised proposal card for approval (e.g. "add that we met on June 14" → updated Note). Keeps the draft→human-approve gate (the human still confirms the LLM's revision) and subsumes `edit field=value` as a deterministic fast path. Thread the instruction text into `normalize`'s source so the email-integrity rule still holds (a revised email must appear in the original message or the instruction). Pairs naturally with the fuzzy-match item above — build both as one conversational-UX pass after the smoke. (Parsing of free-form *intake* messages already works today via the Qwen parse leg; this item is specifically about the *edit/refine* turn.)
+**Post-deploy enhancement — conversational (LLM-mediated) edits (Grant, 2026-06-17). BUILT 2026-06-17 (bot-side, ships on the Spark), not yet deployed / live-smoked.** Today an in-thread correction uses a rigid grammar (`edit field=value`). Let a free-form reply that isn't `yes`/`no`/a literal `edit …` be treated as a natural-language revision instruction: send {current proposal + the instruction} back through local Qwen (`spark.py`, the same parse leg — no Claude, no scrub) and re-render the revised proposal card for approval (e.g. "add that we met on June 14" → updated Note). Keeps the draft→human-approve gate (the human still confirms the LLM's revision) and subsumes `edit field=value` as a deterministic fast path. Thread the instruction text into `normalize`'s source so the email-integrity rule still holds (a revised email must appear in the original message or the instruction). Pairs naturally with the fuzzy-match item above — build both as one conversational-UX pass after the smoke. (Parsing of free-form *intake* messages already works today via the Qwen parse leg; this item is specifically about the *edit/refine* turn.)
+- **As built:** `parse.revise` + `_apply_revision` (offline-testable; the approval-stage `else` branch in `bot.py` routes any non-yes/no/edit reply here). `parse_message` now stashes `_source_text` so revise can re-check email integrity against {instruction + original}; the model's email field is never trusted. No-op revisions are caught via `proposals.same_fields` (re-prompt, not a false "Updated"). **Known v1 limit:** revise edits fields but does not re-run the matcher on a mid-thread firm rename. Tests: `matrix_intake/test_parse.py` (revise merge + email integrity + match-id preservation).
 
 ### Scoped service-credential auth path for automated CRM writers
 *Surfaced 2026-06-17 while deploying the Matrix intake bot. **Decision: defer — the bot uses a dedicated member username/password for now.** The CRM has no API-key/service-token path; its only auth is username+password → JWT. A dedicated **member** login is appropriately scoped against what matters operationally (no admin: can't manage users, reset data, or change settings) and unblocks the live smoke today.*