The earlier behavior never went quiet — it dreamed every night, even when
that meant repeating itself. The 'return None on null delta' rule was a
synthesis-doc invention (the dreamer-design-spec.md I treated as
authoritative is itself LLM-generated) that didn't match the actual
desired UX. Aaron called this out.
The repetition problem the quiet rule was claimed to solve is already
addressed in the retrieve layer:
- LLM-generated queries from the observation signal vary nightly
- MMR diversity prevents within-night cluster lock-in
- NREM bias toward under-processed chunks (low consolidation_count)
ensures fresh material gets selected over recently-replayed material
So select_mode now always returns a mode. NREM is the default. Staleness
still routes to Late REM at 3+ days for cross-domain variety. Journal
entries still route to Early REM.
Bug: the previous order checked the "nothing changed → return None" rule
first, so the spec's "corpus unchanged 3+ days → Late REM (shake things
loose)" branch could never fire. Stasis was permanent — quiet would just
keep returning None forever as long as no new chunks or journals appeared,
regardless of how stale the corpus got.
Fix: check staleness first. Quiet remains the default within the 1-2-day
window the spec implicitly grants for the dreamer to "go quiet rather than
manufacturing novelty." At day 3+, Late REM fires automatically — the
spec's mechanism for breaking out of the silence when the corpus isn't
delivering new material.
Observed symptom that triggered this: dreamer fired 2026-05-21 08:00 and
2026-05-22 08:00, both went quiet. Real cause was no new content (which
is correct quiet behavior for days 1-2), but the bug would have made it
stay quiet indefinitely had we not fixed it before day 3.
Implements `dreamer-design-spec.md` lines 27-74: observe_corpus() returns a
signal vector (new_chunks delta, new_journal_entries, recent_questions over
14-day window, days_since_dream, underprocessed_count derived from the new
consolidation cursor); select_mode() returns one of {nrem, early-rem,
late-rem, lucid} or None per the spec's rules. The None return is the spec's
canonical answer to the repetition problem (line 67) — "dreamer goes quiet
rather than manufacturing novelty."
Standalone for now. Not wired into dream_pipeline yet — that happens in the
retrieve() refactor (task #46). dream.py is unchanged in this commit.
Grounded sources cited in module docstring: Friston Active Inference, sleep
research (Stickgold/Walker/Diekelberg & Born), sharp-wave ripples (Buzsáki).
All three appear in BirdAI-Bibliography.md.
Migration prerequisite (already shipped in the prior commit): consolidation
cursor columns last_consolidated_at + consolidation_count added to
embeddings. Backfill from dream-manifest history is task #49.
aaron