v3.12.0

Changed

• Rules-block diet + drift gate (ADR-0036). The managed memory-rules
  block — paid every session in every project by every wired agent — trimmed
  ES 9,746 → 7,979 chars (−18.1%) / EN 9,484 → 7,743 (−18.4%) (~440
  tokens/session) with zero rules dropped (Trust section byte-identical)
  and two rules rescued from a silently-diverged docs variant ("never
  claim to have persisted" when no vault MCP responds; "memory://… is IDE
  memory, not the vault"). The redundant long-form tool catalogue defers to
  the tool descriptions (canonical since ADR-0035). Three new build gates in
  memory-rules-budget.test.mjs: size budget, 13 load-bearing rule phrases
  per language that must survive any trim, and a drift assertion pinning
  AGENTS.md + both docs install pages to the canonical source (they had
  drifted for months under a "keep in sync" comment). Re-run
  create-obsidian-memory (or --rules) to refresh installed blocks.

• Fixed-cost diet: schema budget gate + compact KG defaults + compressed
  session hook (ADR-0035). The per-session cost every wired agent pays
  before calling anything: tool descriptions + server instructions trimmed
  10,226 → 7,641 chars (−25%, ~−650 tokens/session/agent) — operator prose
  cut, usage guidance and security semantics kept verbatim — and a new
  schema-budget.test.mjs fails the build if schema text regrows past
  8,000 chars (or any single description past 450). vault_relations /
  vault_observations MCP default limit drops 200→50 (filter by
  category/tag/note instead). The SessionStart hook now compresses the
  curated index (100 chars per bullet, word-boundary cut; over-cap prose
  dropped whole) instead of blindly truncating at 4,000: measured on the
  reference vault, −21% chars with +70% more pointers visible (23 → 39 of
  63 — the old cut left 40 entries invisible). Re-run create-obsidian-memory
  to refresh installed hooks.

• Compact search wire format + MCP default limit 20→10 (ADR-0034). The
  mission "make Fable 5 cost like Sonnet 5" (~70% cut) on the memory path:
  vault_hybrid_search / vault_fts_search hits now carry only what an agent
  acts on (path, heading/title, snippet, 5-decimal score) — the
  mostly-null ranking diagnostics (bm25_rank, vector_rank, graph_rank,
  rerank_score, raw bm25, 19-digit mtime_ns, full-precision score) cost
  ~20 tokens/hit and ship only under the new explain: true param /
  --explain CLI flag. The MCP default limit drops 20→10 (resolves the
  ADR-0032 deferral; tool descriptions now teach 3–5 for targeted recall).
  Measured on real-vault recalls: −37.8% wire cost on the default call,
  −60% on a broad query that hit the old cap. bench-tokens gains a wire
  arm that counts the exact compact JSON response (overhead charged, not
  hidden): median wire savings vs whole-note reads = 62% at k=3 / 37% at k=5,
  100% answered — CI-gated via the new --assert-wire-savings 0.30. Breaking
  only for consumers that parsed the diagnostic fields (none in-repo; pass
  explain: true to get them back).

• Managed rules block: "Keep it cheap (tokens)" grew from a one-liner into a
  distilled token-discipline section (ES/EN, ADR-0032) — terse-output rules
  from caveman (no filler/tool-call narration/log dumps; technical terms,
  commands and exact errors always verbatim; plain prose returns for
  security warnings, irreversible confirmations and order-sensitive sequences)
  and the YAGNI ladder from ponytail (reuse → stdlib → platform → installed
  dep → one line → minimal code; never simplify validation, data-loss error
  handling or security; non-trivial logic leaves one runnable check). Plus the
  measured limit guidance: pass a low limit (3–5) on targeted recalls.
  Synced across memory-rules.mjs, AGENTS.md and docs/{es,en}/install.md.

• Memory report notice now carries the compression-safety rule (from
  caveman-compress's safety tests): when condensing notes, compress prose only —
  never drop decisions, gotchas, commands, or exact error strings.

Added

• vault_edit_file now rejects self-paste frontmatter duplication (ADR-0033).
  Found via a vault-hygiene pass: a past corrupt edit had triplicated a project
  note by pasting its own YAML frontmatter (and the rest of the file) back into
  its body, growing it to ~53K tokens undetected until a manual audit. The
  guard is a targeted, deterministic check — if applying an edit's newText
  would increase how many times the note's own frontmatter block occurs, the
  whole call is rejected and the file is left untouched. 3 new tests in
  vault-fs.test.mjs.
• Token-economy benchmark, measured and CI-gated (ADR-0032). The kit's token
  claim ("passage-first reads beat whole-note reads") is now a number, not an
  assertion: bench-tokens / json-bench-tokens compare the top-k
  heading + snippet passages vault_hybrid_search returns against every
  ground-truth note read whole, over a new realistically-sized fixture
  (evals/tokens/, 16 labelled queries). Methodology distilled from the two
  tools evaluated in ADR-0032: caveman's honest control arm (the whole-note arm
  pays zero discovery cost, so savings are a floor; median/min/max/stdev
  reported, ~4 bytes/token estimator on both arms) and ponytail's completeness
  gate (savings only count when every relevant note surfaces in the top-k).
  Measured at k=5: 100% answered, median savings 47%, aggregate 56% — and the
  per-kind breakdown keeps the honest counterpoint visible (small <2 KB notes
  are cheaper read whole). CI gates at --assert-savings 0.40 --assert-answered 0.95; 6 new tests in test_bench_tokens.py also pin that
  a lower k saves strictly more without losing answers.