v0.11.0

Every new lever in this release defaults OFF (backward-compatible): with the
defaults, hook output stays byte-identical to 0.10.0.

Added

• memory.max_content_length is now the single source of truth for the memory
  content cap (#79). The configured cap (default 200) drives both enforcement
  and the "keep content under N chars" guidance in every extraction / digest /
  episode / reflection prompt, plus the content-producing truncations (simba rule add, auto-learned TOOL_RULE, SYSTEM sync rows) — previously these hardcoded
  200 in ~13 places, so raising the cap silently had no effect. Resolved via
  simba.memory.config.resolve_max_content_length(). Set it once
  (simba config set memory.max_content_length <N>, or per-project) and all
  guidance + truncation follow. content is what recall surfaces (not the
  unbounded context), so a larger cap means more tokens per memory and, under
  the recall token budget, can surface fewer memories per turn.

• Conditional guardian re-injection (spec 25; hooks.guardian_signal_gated,
  default off). The CLAUDE.md SIMBA:core block is re-injected every prompt
  (~2k tokens) regardless of need. When enabled, the block is re-injected only on
  the first prompt of a session, after compaction, or when the prior response
  dropped the [✓ rules] signal — reclaiming the per-turn token tax. Fail-open
  (inject when uncertain). A per-session flag written by Stop, read by
  UserPromptSubmit, reset by PreCompact/SessionStart. Plus documented
  gate-graduation discipline (a rule that becomes a deterministic gate leaves the
  CORE block).

• Hierarchical (ancestor-prefix) project recall (spec 26;
  memory.hierarchical_recall + hierarchical_recall_include_global, default
  off). A child cwd (/repo/api) inherits memories scoped to its ancestors
  (/repo) + global, bounded at the git root — the client computes the scope
  chain, the daemon does string-membership across both the vector and BM25 arms;
  projectPath is normalized to a resolved absolute path on store. Dissolves the
  dedup dual-home problem (place a fact once at the right level). Measured
  (A/B on the production stack): a situational lever — near-free at low
  ancestor:child noise ratios, a real recall@k cost at high ratios with sparse
  gold (LoCoMo 9:1 ≈ −0.7pp, longmemeval_s 14.8:1 −4 to −9pp) — so it stays
  default-off and is enabled per-project where the ancestor scope is small.

• Intent-primed doctrine + mandated preflight (spec 28;
  hooks.intent_priming_enabled / preflight_mandate_enabled /
  preflight_mandate_risk_only, default off). UserPromptSubmit classifies the
  prompt against project doctrine triggers (cosine over precomputed embeddings,
  no LLM) and injects the matched doctrine + applicable gates; on a risk-tier
  match it arms a mandate, and PreToolUse then blocks a mutating tool that runs
  without a simba preflight <task> this turn. New simba preflight CLI +
  /preflight endpoint.

• Reasoning-layer verification (spec 27; hooks.engagement_marker_enabled /
  reasoning_verify_enabled, default off). A simba-emitted 🦁☑ ledger of what
  simba surfaced this turn is injected at UserPromptSubmit (every turn, tools or
  not), the agent echoes it, and Stop verifies the echo — observability, not
  self-attestation. reasoning_verify lets Stop/SubagentStop (the latter newly
  wired) doctrine-check the finalized response and block-to-reconsider on a
  violation — the catch for tools-free / mid-reasoning output. On pi only, the
  context event re-injects doctrine before every LLM call and message_end
  annotates a doctrine-violating finalized message (Tier 2; pi context injection
  pending runtime verification).

Changed

• CI hardening. A per-test pytest-timeout (--timeout=120 --timeout-method= thread) plus a job-level timeout-minutes: 20 turn a hung test into a fast
  fail instead of running to the 6-hour Actions ceiling. A global autouse fixture
  forbids real GGUF model loads in the unit suite (reranker fail-opens;
  gguf-marked tests exempt), so CI never reaches Hugging Face.