v0.46.19.0

Dream-cycle synthesis is now fast by default: transcript synthesis runs as a
single validated model call instead of an agent loop that burned 10+ provider
round-trips per transcript, and the child
drain can run several transcripts at once. Six companion fixes make subagent
jobs honest about truncation, failed writes, and provider quirks.

Added

• Oneshot dream synthesis (default). dream.synthesize.mode (default
  oneshot) replaces the per-transcript agentic loop with ONE structured
  completion; pages are validated (slug fences, task shapes, hash suffix,
  exact-match wikilinks) before ANY write, then written programmatically
  through the same put_page executor with deferred embeds. Any validation
  failure falls back to the agentic loop in the same job with a
  fallback_reason operators can read from phase telemetry. Revert dial:
  gbrain config set dream.synthesize.mode agentic. (garrytan#4216)
• Pre-retrieval LINK CANDIDATES manifest. Wikilink targets are resolved
  BEFORE the model call from triage-cached entities/segments (zero-embed:
  basename index + keyword search), so both modes stop burning search turns.
  dream.synthesize.link_manifest (default on). The synthesis prompt now
  also carries the write allow-list explicitly. (garrytan#4216)
• Bounded inline-drain concurrency. dream.synthesize.inline_concurrency
  (default 1, clamp [1,8], PGLite stays serial) runs multiple synthesis
  children at once; phase details now report queue-wait/runtime p50/p95,
  drain wall-clock, per-mode job counts, and a fallback-reason histogram.
  Rate leases stay the provider ceiling — gateway-loop turns now acquire a
  per-turn lease permit (they previously ran unleased), and a lease-full
  child requeues without burning an attempt. (garrytan#4194)
• Structural write accounting. Every subagent job result now carries
  pages_attempted/written/failed (oneshot results also carry
  written_refs); dream/patterns children
  set require_writes so a job whose every write failed goes to the dead
  letter with the first real error instead of reporting success. Phase
  cooldowns are only stamped on a fully-successful run, and a run where every
  child died is an honest phase failure. (garrytan#4217)

Fixed

• Gateway tool loop reports output-cap truncation as max_tokens instead of
  a clean completion; oneshot treats truncated JSON as a fallback, never a
  parse. (garrytan#4088)
• Thinking-by-default Claude 5 models are detected under provider-prefixed
  and bare ids (openrouter:anthropic/claude-*-5, claude-cli:*), so the
  32k output headroom applies wherever it should — and never to
  8k-capped 3.5-era models. (garrytan#4087)
• claude-cli subagent jobs no longer dead-letter when the model repeats a
  tool id across turns: ids are minted locally, and a uniqueness-violation
  backstop covers any provider that repeats ids. (garrytan#4155)
• Gemini 3.x multi-turn tool loops work: per-part provider metadata
  (thoughtSignature) survives the round-trip and crash-replay. (garrytan#4201)
• Crash-safe oneshot recovery: the whole write batch is banked atomically
  before the first write, a retried job finalizes from the write ledger
  instead of re-calling the model, interrupted writes are re-executed, and
  in-batch link edges are replayed — a completed synthesis job can no longer
  silently lose pages. (ship-review + red-team hardening)
• Legacy note: from this release on, a synthesis child whose every write
  failed dead-letters, and dead jobs release their idempotency keys — so the
  next gbrain dream --phase synthesize (or the nightly) retries those
  transcripts automatically. Rows from PRE-upgrade jobs that reported
  success without writing pages stay completed and keep their keys, so
  those transcripts are not retried automatically; any content change to
  the transcript re-keys it and the next run picks it up.

Changed

• gbrain agent logs audit lines now show the synthesis mode and fallback
  reason when present.