Phase 1.1 + 1.2 of the multi-phase plan.

# 1.1 — additive rewrite

Replaces multiplicative tier weighting (`score *= weight`) with additive
bonuses (`score += bonus`). The real-embedding ablation in PR #272
showed multiplicative was structurally bounded — a 1.5×/0.8× swing
(1.875× ratio) let weak-overlap authored content leapfrog strong
primary hits regardless of relevance. Additive bonuses (~±0.005 in
the normal band) can flip close calls but never leapfrog strong
matches.

Promote-only configuration: only authored tiers get a positive bonus;
all received tiers are 0. Trying any negative bonus pushed legitimate
primary hits on `received_content` queries below distractors. The
product intent is "prefer authored on close calls," not "suppress
received" — promote-only gives the former without the latter.

Floor-ratio gate retained (default 0.85). Real embedders give
non-trivial cross-query vector similarity; without the gate,
authored content from unrelated queries leaks into the candidate
pool and gets boosted past legitimate primaries.

Files:
- tier-weights.ts: `tierBonus` / `buildTierBonusFn` (additive).
  `tierMultiplier` / `buildTierWeightFn` re-exported as deprecated
  aliases for back-compat.
- rrf.ts: applies bonus additively, NEGATIVE_INFINITY sentinel for
  hidden, 0.85 floor-ratio gate.

# 1.2 — flip default-on

Phase 1.1 cleared the eval bar:

  user_behavior MRR       0.667 → 1.000  (preserved)
  received_content MRR    1.000 → 0.833  (real embeddings)
                                  → 0.583  (hash-trick floor)
  aggregate MRR primary   0.857 → 0.929  (above pure-RRF baseline)

Files:
- Migration 044: ALTER DEFAULT true + backfill existing rows.
- parseSettingsRow / in-memory + CRDB upsert / route GET — all default
  flipped to true.

Tests (98 pass, 70 turbo tasks green):
- tier-weights.test.ts: 19 cases updated for additive semantics, all 3
  calibrations, override composition, back-compat aliases.
- rrf.test.ts: new "weak-match doesn't leapfrog" case; existing cases
  reformulated for additive bonus.
- tier-ablation-eval bars tightened: received_content ≥ 0.55
  (hash-trick), ≥ 0.75 (real embeddings).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Five findings, all valid:

1. JSDoc had "Number. NEGATIVE_INFINITY" split across two lines —
   reads awkwardly in generated docs. Joined.

2. Inclusion-semantics drift: the post-bonus filter was `rrfScore > 0`,
   which silently dropped pages with sufficiently-negative bonuses
   alongside the intended NEGATIVE_INFINITY-sentinel drops. Tightened
   the filter to only remove the sentinel; negative bonuses now reorder
   without changing inclusion. Documented in the TierWeightFn JSDoc.

3. Migration 044's comment claimed "only rows that were never explicitly
   toggled" get backfilled, but the SQL unconditionally flips all
   tier_weighting=false rows. We don't have a "set by user" audit
   column to distinguish defaults from opt-outs, so the honest fix is
   to update the comment — clarifies that this IS an unconditional
   opt-in. Notes that a future audit column could preserve opt-outs
   if it becomes important.

4. The "doesn't leapfrog" test in rrf.test.ts had exploratory scratch
   notes including a "PR #_" placeholder and a self-contradicting "Wait
   — additive DOES flip this" line. Replaced with a clean explanation
   of the fixture being asserted, the actual rank/score numbers, and
   the load-bearing role of the 0.85 floor-ratio gate.

5. tier-weights.ts docstring said "rank-1 vs rank-2 RRF diff is ~0.001"
   but the table below shows 0.0164 vs 0.0161 = 0.0003. Corrected
   to 0.0003.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>