rmerk turns Mike into a Minnesota med-mal records engine

Plan for a "Medical Malpractice Case" project template that scaffolds
the standard subfolder structure on project creation and registers
reusable tabular-review column schemas (chronology, bills, MAR,
transfusion log, vitals, labs, imaging index, red-flag scan) the user
can instantiate once documents are uploaded.

Three new docs and a substantive revision to the templates plan, all
captured in `docs/`:

- `RESEARCH_mn_med_mal_law.md/.pdf` - validates the med-mal-case template
  against Minnesota primary sources (statutes, MN Supreme Court opinions,
  Rules of Civil Procedure). Corrects two inherited errors: § 145.682(4)'s
  180-day clock runs from Rule 26.04(a) discovery commencement (not
  summons service), and the MN prima facie negligence test is the
  3-element Plutshack/Smith v. Knowles formulation (not the 5-element
  informed-consent test). Ends with a 10-item "Plan deltas" appendix.
- `RESEARCH_legal_rag.md` - surveys legal RAG approaches across case-law,
  contracts, e-discovery, and med-mal lanes; recommends structured
  extraction with page+bbox citations over vector RAG for med-mal records.
- `PLAN_med_mal_extraction_pipeline.md` - Mike-specific build plan derived
  from the RAG research: page-by-page multimodal extraction → JSON event
  log in Postgres → deterministic red-flag rules → page-window retrieval
  for chat. Out-of-scope: embeddings, GraphRAG, cross-case search.
- `PLAN_med_mal_templates.md` - applies all ten plan deltas: subfolder
  taxonomy expanded 8→12+nested (adds expert-affidavits-145682,
  hipaa-authorizations, discovery, collections-liens-subrogation,
  pre-retainer-investigation; renames bills→bills-and-eobs); three new
  tabular schemas (provider-defendant-map, causation-chain,
  expert-opinions-145682); bills columns split per § 548.251; red-flag
  scan gains § 145.682 affidavit link + temporal-proximity anchor; new
  "Cited authority" section consolidates statutes/rules/cases for future
  revision audits. v1 schema change stays at single `template_id` column;
  v1.1 follow-up note flags `key_dates jsonb null` for deadline tracking.

Applies the same MN-law validation pass to PLAN_med_mal_extraction_pipeline.md
that just landed on PLAN_med_mal_templates.md, surfacing ten substantive
deltas plus an architectural tension between the event log and tabular review
schemas. Net effect: the multi-day Phase 2 build no longer inherits the
breach-only red-flag bias or the missing MN-specific privacy/work-product
defenses that the original plan carried.

Schema changes (folded into the inline SQL):
- document_events gains provider_role (Plutshack role-specific SOC),
  episode_of_care (chronology clustering), privacy_class (§ 144.293 /
  § 145.64 / 42 CFR Part 2 segregation), and key_date_role (feeds Phase 4
  key_dates jsonb deadline-tracking widget). medications jsonb shape
  expanded with ordered_at / administered_at / allergy_conflict_flag /
  weight_based_dose_check_passed for the Mulder rule (Reinhardt).
- document_red_flags gains supports_element (Plutshack/Smith 4-cut) and
  awaits_expert_affidavit (closes loop with § 145.682(4)(a) checklist).
- RLS clause added to filter peer_review_145_64 rows from default queries.

Red-flag library rebalanced from 5 breach-only rules to 6 rules tagged
across the duty/breach/causation framework, including a new
temporal_anchor_causation rule that surfaces tight temporal anchors without
asserting causation (Plutshack/Smith still require expert testimony).

§Defenses expanded with four MN-specific policies:
- § 145.64 peer-review hard-refuse: extraction halts entirely on
  peer-review-marked documents; no events written. This is the strictest
  policy in the codebase.
- § 144.293 mental-health redaction-by-default until project-level toggle
  confirms the heightened authorization is on file.
- Rule 408 settlement-comms caveat: extracted but narrative prefixed with
  [Rule 408 - inadmissible to prove liability].
- Rule 26(b)(3) work-product opt-out: consulting-expert notes NOT extracted
  by default; project-level opt-in only.

§Out of scope grew with two permanent architectural separations:
- Causation-chain reasoning reserved for builtin-causation-chain tabular
  review (Plutshack/Smith expert testimony requirement; Dickhoff
  loss-of-chance fact-intensity).
- Provider-defendant entity resolution reserved for
  builtin-provider-defendant-map (Popovich two-factor test requires
  outside-the-record evidence - Rock v. Abdullah reliance-element limit).

Adds a Plan deltas appendix and Cited authority section mirroring the
research doc's format so future revisions can re-check against the same
source set. PDF re-rendered via reconstructed /tmp/jenn-mike/md_to_pdf.py
(markdown + weasyprint).

Phase 1 of the templates → extraction-pipeline path. Selecting a template
when creating a project scaffolds the 12-folder med-mal taxonomy + nested
imaging folder, and exposes a "Recommended tabular reviews" strip on the
project page with one-click buttons for the 11 schemas validated against
Minnesota primary sources (Plutshack/Smith, § 145.682, § 548.251,
§ 144.293, Popovich, Dickhoff, Flom, Reinhardt, Cornfeldt).

Backend:
- New table column: projects.template_id (text null) - applied directly to
  prod (`qkfcrsrtualqdmqqexpf`) via Supabase MCP apply_migration after
  branch creation returned PaymentRequiredException on the free plan.
  Folded into schema.sql + backend/migrations/0001_projects_template_id.sql.
- backend/src/lib/templateIds.ts - shared string-literal unions for
  ProjectTemplateId and TabularSchemaId.
- backend/src/lib/builtinProjectTemplates.ts - registry exporting the
  med-mal-case template with its 12 subfolders and 11 recommended schema ids.
- backend/src/routes/projects.ts - POST /projects extended to parse
  optional template_id, validate against registry (400 on unknown), and
  batch-create subfolders in two passes (top-level UUIDs returned by
  Batch A resolve the array-index parent refs in Batch B). On any error
  the project row is deleted via CASCADE so the operation is atomic from
  the user's perspective.

Frontend:
- frontend/src/app/components/tabular/templateIds.ts - mirror of backend
  template/schema IDs.
- frontend/src/app/components/tabular/builtinTabularSchemas.ts - the only
  place column specs for built-in schemas live (11 schemas, ~120 columns
  total) with prompts tuned for Mayo/Epic ebook formatting. Each schema's
  description names the prima facie element or filing requirement it
  supports.
- frontend/src/app/components/projects/builtinProjectTemplates.ts -
  frontend mirror of the template registry (id, name, description,
  recommendedSchemaIds only; backend owns subfolders).
- frontend/src/app/lib/mikeApi.ts - createProject signature gains
  optional template_id arg.
- frontend/src/app/components/shared/types.ts - MikeProject gains
  template_id and subfolders fields.
- frontend/src/app/components/projects/NewProjectModal.tsx - adds a
  Template dropdown between the CM number field and the attribute pills,
  using the same Check-icon-on-select pattern as AddNewTRModal's workflow
  picker.
- frontend/src/app/components/projects/ProjectPage.tsx - on the reviews
  tab, renders a recommended-reviews strip when project.template_id is
  set. Each pill opens AddNewTRModal with the schema's title and full
  columns_config pre-populated.
- frontend/src/app/components/tabular/AddNewTRModal.tsx - accepts new
  optional initialTitle and initialColumnsConfig props that the
  recommended strip uses to seed the modal.

CLAUDE.md: documents the migrations workflow + flags the Supabase
branching limitation on the free plan so future migrations don't reach
for create_branch first.

Closes the documentation gap surfaced after Phase 1 shipped: the
per-phase plans (PLAN_med_mal_templates.md, PLAN_med_mal_extraction_pipeline.md)
were already committed, but the meta-orchestration that sequences them
lived only in a local Claude planning file. Future contributors (and
resumed sessions) now see the whole arc in one repo-resident document.

The roadmap captures:
- Why Phase 0 (research-validation pass) precedes Phase 2 - cheap
  insurance that produced 10 deltas to the extraction-pipeline plan
  including a missing causation rule, missing § 145.64 peer-review
  hard-refuse, and missing Plutshack/Smith 4-cut tagging.
- Why Phase 1 (templates) precedes Phase 2 (extraction) - exercises the
  Supabase-branch + schema-migration toolchain on a 1-column reversible
  change before the 3-table change.
- Four hidden dependencies surfaced from codebase exploration that the
  per-phase plans don't address: missing backend/migrations/ directory,
  text-only LLM provider abstraction, no bbox-extraction primitive,
  free-plan Supabase branching limit.
- Phase status (0 + 1 shipped at 3a97315 and 73faac0; 2/3/4 pending).
- Concrete Phase 2 first steps + the multimodal/bbox sub-tasks hidden
  inside the multi-day estimate.
- Phase 3 architectural decision (event log stays narrow; per-schema
  extractors consult it rather than expanding the log).
- Phase 4 v1.1 deadline-tracking surface tied to Phase 2's
  document_events.key_date_role column.
- Cross-phase risk register, out-of-scope list with permanent
  architectural separations (causation-chain reasoning,
  provider-defendant entity resolution).

PDF rendered via reconstructed /tmp/jenn-mike/md_to_pdf.py.

Postgres (`0002`-`0005`), `patch_document_extraction_run` with GRANTs, per-page Claude JSON extraction, raster + vision for empty text layers, §145.64 vision-page peer-review prescan (compliance gate), optional queue mode (`EXTRACTION_ASYNC_MODE=queue`), REST + UI + chat tools, Vitest and Supertest coverage (403/404/409), backend CI workflow.

Closes the §145.64 compliance gap: scanned pages with peer-review marker phrases visible only in the raster are now detected and halt the run before any event-extraction call.

Follow-up: apply `0005_extraction_async_jobs_document_index.sql` on each Supabase environment and verify advisor `unindexed_foreign_keys` clears.

Phase 3 of the med-mal records platform. Surfaces the Phase 2
document_events log as a fast, citation-anchored chronology view -
satisfying the roadmap §Phase 3 verification gate of "+ Medical
Chronology on extracted doc loads in <2s, no LLM call."

Changes:
- New route /projects/[id]/timeline/[docId] rendering events as a
  sortable table with row-click bbox sync against the embedded PDF
  preview. Reads the existing GET /extraction/:documentId/events;
  zero backend changes.
- Widen MedMalDocumentEvent type to expose provider, provider_role,
  event_time, episode_of_care, key_date_role, event_date_text.
- + Medical Chronology on ProjectPage routes to Timeline when ≥1 PDF
  is ready: single PDF goes straight, multiple opens a picker, zero
  PDFs falls back to the existing tabular-create modal.
- docs/PLAN_med_mal_phase3_integration.md documents the row-cardinality
  finding that ruled out the originally-planned Rail B (cells-from-event-log
  for MAR / vitals / etc.) - every event-log-backed schema is per-encounter,
  which doesn't fit the 1-row-per-doc tabular_cells model. Future MAR /
  vitals / labs views are deferred to Phase 3.5 as their own Timeline-style
  surfaces.

Verification: builds clean on backend + frontend; lint clean for new
file. Browser-clicked e2e is left to the user - the gate
("<2s, no LLM traffic") needs a running stack + an extracted doc to
exercise, neither of which the agent had access to during this PR.

Standalone tsx script piping synthetic page-text through MedGemma 27B-IT via
LM Studio's OpenAI-compatible endpoint, mirroring the medMalExtractor prompt
shape. Validates clinical-field quality and JSON/shape stability ahead of
deciding whether to wire it in as a stage-2 normalization model.