feat(t-paliad-190): mig 090 — one-time fuzzy-match backfill

Phase 3 Slice 10 Step I (design §3.I + m's Q10 ruling). Binds legacy
paliad.deadlines.rule_id to deadline_rules.id via priority-ordered
fuzzy matching; ambiguous + no-match rows log to the orphan staging
table (mig 089).

Matching strategies (highest priority first; first unique hit wins):

  1. rule_code_and_tail — title's leading citation token AND its
     post-separator name fragment match a rule. Handles
     "RoP.023 — Klageerwiderung" where the bare code matches 2 rules
     (DE Klageerwiderung + EN Statement of Defence); the tail picks
     the right one.

  2. rule_code only — bare rule_code from the title prefix. Handles
     "RoP.029.a — Replik" where RoP.029.a maps to a single rule
     regardless of suffix (the title's "Replik" doesn't match the
     rule's actual name but the code is unique).

  3. name_exact — full title equals rule.name or rule.name_en
     (LOWER). Catches "Antrag auf Schadensbemessung" (1 unique
     rule); ambiguous for shared names like Klageerwiderung (8
     candidates).

  4. concept_alias — title appears in deadline_concepts.aliases.
     Thin coverage today; Slice 12 orphan-seed will populate it.

Per-deadline aggregation:
  - Strategy with n_candidates = 1 wins. Priority chain rule_code_and_tail
    > rule_code > name_exact > concept_alias.
  - Ambiguous (≥2 across all strategies) → orphan reason='ambiguous'
    with the full candidate_rule_ids list.
  - 0 candidates → orphan reason='no_match'.

Predicted production outcome (verified via supabase MCP pre-write):
  - 3 of 25 deadlines (12%) get a unique match:
      "RoP.023 — Klageerwiderung"   via rule_code_and_tail
      "RoP.029.a — Replik"          via rule_code
      "Antrag auf Schadensbemessung" via name_exact
  - 15 of 25 deadlines (60%) → orphan reason='ambiguous' (common
    titles like Klageerwiderung × 4, Duplik × 4, Replik × 4 across
    multiple proceedings).
  - 7 of 25 deadlines (28%) → orphan reason='no_match' (free-text
    titles like "Call me", "Schutzschrift", "Validierungsfrist EP→DE",
    "Schriftsatz nach R.262 (Klageerwiderung)").

The 60% target the design § hinted at is unachievable on today's
corpus because all 11 projects have proceeding_type_id IS NULL post-
Slice-5 (the fristenrechner-side rebinding hasn't happened on
production data yet) — proceeding-narrowing would cut the
Klageerwiderung / Duplik / Replik ambiguity, but the column isn't
populated. The orphan-review UI in Slice 11 is the real path to
binding the long tail.

Defensive backup: paliad.deadlines_pre_089 snapshot taken before any
UPDATE. Down-migration restores rule_id from the snapshot + drops
unresolved orphan rows (resolved rows survive a rollback — those are
legal-review work that shouldn't disappear on a code revert).

Idempotency: WHERE rule_id IS NULL on the UPDATE; orphan INSERT
skips rows that already have an unresolved orphan entry. Re-running
on the same corpus produces no new rows.

Hard assertion: every NULL-rule_id deadline (with project) is either
resolved post-mig OR has an unresolved orphan row. RAISE EXCEPTION on
any unaccounted row — fails the migration loudly rather than
silently leaving a deadline un-matched + un-orphaned.

Audit-reason wrapper set; the mig 079 deadline_rules audit trigger
doesn't fire here (UPDATEs touch paliad.deadlines, not deadline_rules),
but the wrapper is the standard pattern.

internal/db/migrations/090_backfill_deadline_rule_id.down.sql

@@ -0,0 +1,30 @@
+-- t-paliad-190 down — reverses 090_backfill_deadline_rule_id.up.sql.
+--
+-- Restores rule_id values from the pre-mig snapshot (every deadline
+-- that mig 090 touched had rule_id IS NULL originally, so restoring
+-- means setting rule_id back to NULL on every row that survived the
+-- backfill). Drops the orphan rows mig 090 wrote (resolved rows stay
+-- — those represent legal-review work that shouldn't disappear on
+-- a code rollback) and drops the backup table.
+--
+-- This is a defensive rollback path; the migration itself is one-time
+-- + idempotent, so re-running 090 after a down + up is safe.
+
+SELECT set_config(
+    'paliad.audit_reason',
+    'rollback 090: NULL rule_id on deadlines mig 090 touched + drop pre-089 backup',
+    true);
+
+-- Restore rule_id = NULL on every deadline mig 090 may have written.
+-- We use the backup table as the authoritative "before" snapshot.
+UPDATE paliad.deadlines d
+   SET rule_id = b.rule_id
+  FROM paliad.deadlines_pre_089 b
+ WHERE d.id = b.id;
+
+-- Drop the unresolved orphan rows mig 090 wrote. Resolved rows stay —
+-- a legal-review hand-link is real work that survives a code rollback.
+DELETE FROM paliad.deadline_rule_backfill_orphans
+ WHERE resolved_at IS NULL;
+
+DROP TABLE IF EXISTS paliad.deadlines_pre_089;

internal/db/migrations/090_backfill_deadline_rule_id.up.sql

@@ -0,0 +1,320 @@
+-- t-paliad-190 / Fristen Phase 3 Slice 10 — one-time fuzzy-match
+-- backfill of paliad.deadlines.rule_id per design §3.I + m's Q10
+-- ruling. Restores SmartTimeline's "anchor real deadlines into
+-- projection" affordance on legacy data (1 of 26 deadlines currently
+-- has rule_id populated; the SmartTimeline anchor flow needs the FK
+-- to thread predicted dates off actuals).
+--
+-- Matching strategies (in priority order; first unique hit wins):
+--
+--   1. rule_code-prefix extraction from title. Titles like
+--      "RoP.023 — Klageerwiderung" carry the rule citation in the
+--      prefix; we extract the leading citation token and JOIN on
+--      deadline_rules.rule_code = extracted. When the rule_code
+--      resolves to multiple rules (e.g. RoP.023 → 2 rules — DE
+--      Klageerwiderung + EN Statement of Defence), the remaining
+--      title fragment narrows by name ILIKE.
+--
+--   2. exact title match against rule.name OR rule.name_en (LOWER).
+--      Mostly hits common Pipeline-A names ("Antrag auf
+--      Schadensbemessung" → 1 unique rule); ambiguous for shared
+--      names like "Klageerwiderung" (8 rules across proceedings).
+--
+--   3. deadline_concepts.aliases match. Each concept carries a
+--      text[] of canonical aliases; if LOWER(d.title) is in the
+--      aliases array, we pick the rules with that concept_id. Today
+--      the alias coverage is thin (no aliases for "Schutzschrift"
+--      etc.), but the strategy is shaped so a future seed lights
+--      it up.
+--
+-- For each deadline, we collect all candidates across the three
+-- strategies, dedupe by rule.id, and:
+--   - exactly 1 candidate → UPDATE rule_id (matched).
+--   - 0 candidates → orphan with reason='no_match'.
+--   - ≥2 candidates → orphan with reason='ambiguous', candidate_rule_ids
+--     populated so a legal-review pass can hand-pick.
+--
+-- Per-project narrowing by proceeding_type_id is the design's primary
+-- discriminator. In the live corpus today all 11 projects have
+-- proceeding_type_id IS NULL (Slice 5 retired litigation codes from
+-- project-binding; the fristenrechner-side rebinding hasn't happened),
+-- so this slice can't use proceeding-narrowing on production data.
+-- The CTE still includes the predicate so the migration self-tunes
+-- the moment proceeding_type_id starts getting populated.
+--
+-- Defensive backup: paliad.deadlines is snapshotted to
+-- paliad.deadlines_pre_089 before the UPDATE so an operator can
+-- restore individual rule_id values if a hand-link goes wrong post
+-- mig. The table is dropped in the down-migration; Slice 11 (rule
+-- editor) can drop it once orphan resolution finishes in prod.
+--
+-- Idempotency: WHERE d.rule_id IS NULL on the UPDATE; the orphan
+-- INSERT uses ON CONFLICT DO NOTHING via a NOT EXISTS guard (no
+-- unique constraint on deadline_id alone — a deadline may legitimately
+-- get re-orphaned after a resolution rollback; but re-running 090 on
+-- the same corpus must not duplicate orphan rows for unresolved
+-- deadlines).
+--
+-- Hard assertion at end: SUM(matched) + SUM(orphans for current
+-- unresolved deadlines) ≥ COUNT(deadlines processed). Strict equality
+-- doesn't hold cleanly on a re-run (the orphan table may already
+-- carry prior rows from a partial run), so the assertion is "at
+-- least one row exists per unresolved deadline".
+
+SELECT set_config(
+    'paliad.audit_reason',
+    'mig 090: one-time fuzzy-match backfill of deadlines.rule_id per design §3.I / Q10',
+    true);
+
+-- =============================================================================
+-- 1. Defensive backup before any UPDATE.
+-- =============================================================================
+
+CREATE TABLE IF NOT EXISTS paliad.deadlines_pre_089 AS
+SELECT id, project_id, title, rule_id, rule_code, status, due_date,
+       completed_at, created_at, updated_at
+  FROM paliad.deadlines
+ WHERE rule_id IS NULL
+   AND project_id IS NOT NULL;
+
+COMMENT ON TABLE paliad.deadlines_pre_089 IS
+    'Snapshot of paliad.deadlines (id, rule_id-relevant columns) taken '
+    'before mig 090 ran the fuzzy-match backfill. Lets an operator '
+    'restore individual rule_id values if a hand-link goes wrong. '
+    'Slice 11 (rule editor) drops this once orphan resolution finishes.';
+
+-- =============================================================================
+-- 2. Build the candidate set in a temp table so the per-deadline
+--    aggregation + UPDATE + orphan INSERT can share the work without
+--    re-evaluating the matchers.
+-- =============================================================================
+
+CREATE TEMP TABLE _mig_090_candidates ON COMMIT DROP AS
+WITH targets AS (
+    -- Every NULL-rule_id deadline still bound to a project. project_id
+    -- is required because we want at least the SmartTimeline anchor
+    -- flow to work; un-bound deadlines (rare) are out of scope.
+    SELECT d.id            AS deadline_id,
+           d.title         AS title,
+           d.project_id,
+           p.proceeding_type_id,
+           -- Extract a leading citation token like "RoP.023" or
+           -- "R.49" from the title. Captures the rule_code prefix
+           -- on titles that carry one ("RoP.023 — Klageerwiderung");
+           -- NULL on plain titles.
+           NULLIF(regexp_replace(d.title, '^\s*((?:RoP|R|Art|§)\.?\s*[0-9]+(?:\.[a-z0-9]+)*)\s*(?:[—–-].*)?$', '\1'), d.title) AS code_token,
+           -- Strip the leading citation + separator to surface the
+           -- title's name fragment. "RoP.023 — Klageerwiderung" →
+           -- "Klageerwiderung"; "RoP.029.a" (no suffix) → ""; plain
+           -- "Klageerwiderung" → "Klageerwiderung" unchanged.
+           NULLIF(trim(regexp_replace(d.title, '^\s*(?:RoP|R|Art|§)\.?\s*[0-9]+(?:\.[a-z0-9]+)*\s*[—–-]?\s*', '')), '') AS title_tail
+      FROM paliad.deadlines d
+      LEFT JOIN paliad.projects p ON p.id = d.project_id
+     WHERE d.rule_id IS NULL
+       AND d.project_id IS NOT NULL
+),
+by_code_and_tail AS (
+    -- Strategy 1a (narrowest): rule_code AND name (DE or EN) matches
+    -- the title's tail fragment. Handles "RoP.023 — Klageerwiderung"
+    -- where the bare code matches 2 rules (DE Klageerwiderung +
+    -- EN Statement of Defence); the tail picks the DE one.
+    SELECT t.deadline_id, dr.id AS rule_id, 'rule_code_and_tail' AS strategy
+      FROM targets t
+      JOIN paliad.deadline_rules dr
+        ON dr.rule_code = trim(t.code_token)
+       AND dr.is_active = true
+       AND (LOWER(dr.name)    = LOWER(t.title_tail)
+            OR LOWER(dr.name_en) = LOWER(t.title_tail))
+     WHERE t.code_token IS NOT NULL
+       AND t.title_tail IS NOT NULL
+),
+by_code AS (
+    -- Strategy 1b: rule_code prefix only. Handles bare-code titles
+    -- ("RoP.029.a" maps to 1 unique rule regardless of suffix) and
+    -- the fallback when by_code_and_tail returns 0 (suffix doesn't
+    -- match — e.g. "RoP.029.a — Replik" where the suffix "Replik"
+    -- doesn't appear in any RoP.029.a rule's name; pick the
+    -- code-only match anyway).
+    SELECT t.deadline_id, dr.id AS rule_id, 'rule_code' AS strategy
+      FROM targets t
+      JOIN paliad.deadline_rules dr
+        ON dr.rule_code = trim(t.code_token)
+       AND dr.is_active = true
+     WHERE t.code_token IS NOT NULL
+),
+by_name AS (
+    -- Strategy 2: exact title match against rule.name or rule.name_en.
+    -- The widest matcher; for shared names like "Klageerwiderung"
+    -- (8 rules across proceedings) this is ambiguous, but for
+    -- unique titles like "Antrag auf Schadensbemessung" (1 rule) it
+    -- nails the match.
+    SELECT t.deadline_id, dr.id AS rule_id, 'name_exact' AS strategy
+      FROM targets t
+      JOIN paliad.deadline_rules dr
+        ON (LOWER(dr.name) = LOWER(t.title)
+            OR LOWER(dr.name_en) = LOWER(t.title))
+       AND dr.is_active = true
+),
+by_alias AS (
+    -- Strategy 3: concept aliases. deadline_concepts.aliases is a
+    -- text[] of canonical synonyms; if the deadline title appears
+    -- in that array, every active rule on the concept is a candidate.
+    -- Today's alias coverage is thin (the seed for Slice 12 is the
+    -- expected source of new aliases), but the strategy is in place
+    -- so future seeds light it up without a migration.
+    SELECT t.deadline_id, dr.id AS rule_id, 'concept_alias' AS strategy
+      FROM targets t
+      JOIN paliad.deadline_concepts dc
+        ON LOWER(t.title) = ANY(SELECT LOWER(a) FROM unnest(dc.aliases) a)
+      JOIN paliad.deadline_rules dr
+        ON dr.concept_id = dc.id
+       AND dr.is_active = true
+)
+SELECT deadline_id, rule_id, strategy
+  FROM by_code_and_tail
+ UNION
+SELECT deadline_id, rule_id, strategy
+  FROM by_code
+ UNION
+SELECT deadline_id, rule_id, strategy
+  FROM by_name
+ UNION
+SELECT deadline_id, rule_id, strategy
+  FROM by_alias;
+
+-- =============================================================================
+-- 3. Aggregate per-deadline candidate counts by strategy + pick the
+--    narrowest-unique-match per deadline. Strategy priority (narrowest
+--    first): rule_code_and_tail > rule_code > name_exact > concept_alias.
+--    A deadline's "chosen" rule comes from the highest-priority strategy
+--    that yields exactly 1 distinct candidate.
+-- =============================================================================
+
+CREATE TEMP TABLE _mig_090_strategy_counts ON COMMIT DROP AS
+SELECT deadline_id,
+       strategy,
+       count(DISTINCT rule_id) AS n,
+       MIN(rule_id::text) AS first_rule_text
+  FROM _mig_090_candidates
+ GROUP BY deadline_id, strategy;
+
+CREATE TEMP TABLE _mig_090_chosen ON COMMIT DROP AS
+SELECT DISTINCT ON (deadline_id)
+       deadline_id,
+       first_rule_text::uuid AS rule_id,
+       strategy AS chosen_strategy
+  FROM _mig_090_strategy_counts
+ WHERE n = 1
+ ORDER BY deadline_id,
+          CASE strategy
+              WHEN 'rule_code_and_tail' THEN 1
+              WHEN 'rule_code'          THEN 2
+              WHEN 'name_exact'         THEN 3
+              WHEN 'concept_alias'      THEN 4
+              ELSE 5
+          END;
+
+-- "Aggregated" carries the widest candidate set for orphan logging
+-- (an editor reviewing an orphan wants to see EVERY plausible rule,
+-- not just the narrowest-strategy result).
+CREATE TEMP TABLE _mig_090_aggregated ON COMMIT DROP AS
+SELECT c.deadline_id,
+       count(DISTINCT c.rule_id) AS n_candidates,
+       array_agg(DISTINCT c.rule_id) AS all_rule_ids
+  FROM _mig_090_candidates c
+ GROUP BY c.deadline_id;
+
+-- =============================================================================
+-- 4. UPDATE deadlines.rule_id for the chosen set (narrowest-unique match).
+-- =============================================================================
+
+UPDATE paliad.deadlines d
+   SET rule_id = c.rule_id
+  FROM _mig_090_chosen c
+ WHERE d.id = c.deadline_id
+   AND d.rule_id IS NULL;
+
+-- =============================================================================
+-- 5. Log every deadline that didn't get a unique match as an orphan.
+--    Skip rows that already have a non-resolved orphan entry (re-run
+--    guard) — the existing entry is the source-of-truth until the
+--    admin UI flips resolved_at.
+-- =============================================================================
+
+INSERT INTO paliad.deadline_rule_backfill_orphans
+    (deadline_id, title, project_id, proceeding_code, reason,
+     candidate_count, candidate_rule_ids)
+SELECT t.deadline_id,
+       t.title,
+       t.project_id,
+       pt.code AS proceeding_code,
+       CASE
+           WHEN a.n_candidates IS NULL OR a.n_candidates = 0 THEN 'no_match'
+           WHEN a.n_candidates > 1                            THEN 'ambiguous'
+       END AS reason,
+       COALESCE(a.n_candidates, 0),
+       COALESCE(a.all_rule_ids, ARRAY[]::uuid[])
+  FROM (
+        SELECT d.id AS deadline_id, d.title, d.project_id, p.proceeding_type_id
+          FROM paliad.deadlines d
+          LEFT JOIN paliad.projects p ON p.id = d.project_id
+         WHERE d.rule_id IS NULL
+           AND d.project_id IS NOT NULL
+       ) t
+  LEFT JOIN _mig_090_aggregated a ON a.deadline_id = t.deadline_id
+  LEFT JOIN paliad.proceeding_types pt ON pt.id = t.proceeding_type_id
+ WHERE NOT EXISTS (
+       SELECT 1
+         FROM paliad.deadline_rule_backfill_orphans o
+        WHERE o.deadline_id = t.deadline_id
+          AND o.resolved_at IS NULL
+ );
+
+-- =============================================================================
+-- 6. Hard assertion: every NULL-rule_id deadline (with project) is
+--    either resolved (rule_id IS NOT NULL post-mig) or carries an
+--    unresolved orphan row.
+-- =============================================================================
+
+DO $$
+DECLARE
+    n_processed   int;
+    n_matched     int;
+    n_orphaned    int;
+    n_unaccounted int;
+BEGIN
+    SELECT count(*) INTO n_processed
+      FROM paliad.deadlines
+     WHERE project_id IS NOT NULL
+       AND (rule_id IS NOT NULL OR EXISTS (
+            SELECT 1 FROM paliad.deadline_rule_backfill_orphans o
+             WHERE o.deadline_id = paliad.deadlines.id
+       ));
+
+    SELECT count(*) INTO n_matched
+      FROM paliad.deadlines d
+      JOIN paliad.deadlines_pre_089 b ON b.id = d.id
+     WHERE d.rule_id IS NOT NULL;
+
+    SELECT count(DISTINCT deadline_id) INTO n_orphaned
+      FROM paliad.deadline_rule_backfill_orphans
+     WHERE resolved_at IS NULL;
+
+    SELECT count(*) INTO n_unaccounted
+      FROM paliad.deadlines d
+     WHERE d.rule_id IS NULL
+       AND d.project_id IS NOT NULL
+       AND NOT EXISTS (
+           SELECT 1 FROM paliad.deadline_rule_backfill_orphans o
+            WHERE o.deadline_id = d.id
+       );
+
+    RAISE NOTICE 'mig 090: processed=% matched=% orphaned=% unaccounted=%',
+                 n_processed, n_matched, n_orphaned, n_unaccounted;
+
+    IF n_unaccounted > 0 THEN
+        RAISE EXCEPTION 'mig 090: % deadlines have rule_id IS NULL and no orphan row — '
+                        'matcher missed them. Investigate the candidate query.',
+                        n_unaccounted;
+    END IF;
+END $$;