Clarify rumination deepening guard

notes/fresh-workspace-handoff.md

@@ -41,7 +41,7 @@ Git history and smoke artifacts carry routine proof. `state/evidence.jsonl` shou
 - The old memory-organ script family, legacy moderation-state template, and legacy moderation context exporter have been deleted. Do not revive their cleanup, agency, dream, vector-incubation, or mirror-repair behavior as compatibility scaffolding.
 - `scripts/run-void-memory-maintenance.ps1` is the first replacement memory-maintenance organ. It builds a bounded context from typed memories/incubation/candidates/runtime, loads `prompts/void-memory-maintenance.md`, and applies only allowed memory/incubation/candidate typed operation proposals through `scripts/void-self-state.mjs`. In sleep mode, memory pressure sets `sleepDirective.forceDistillation`; a real model pass returning no operations under pressure fails visibly.
 - Rumination does not write durable memory. It can only emit `record_short_term_memory`; sleep maintenance must promote short-term records with `apply_memory_distillation`, merge them into incubation and prune them, or prune them outright. A real sleep pass that leaves short-term records behind fails.
-- Short-term rumination memory is clustered at the typed service boundary now. When a new `record_short_term_memory` operation repeats the same target/topic pressure, the service keeps one provisional memory, merges anchors/tags, preserves a stable source id for sleep, and updates the thought instead of stacking duplicates. `npm run moderation:short-term-clustering:fixture` proves the guard.
+- Short-term rumination memory is clustered at the typed service boundary now. When a new `record_short_term_memory` operation repeats the same target/topic pressure, the service keeps one provisional memory, merges anchors/tags, preserves a stable source id for sleep, and updates the thought instead of stacking duplicates. This is meant to protect deepening, not suppress it: low-saturation or newly anchored thoughts should keep developing inside the cluster until they stop changing. `npm run moderation:short-term-clustering:fixture` proves the guard.
 - Long-term memory is durable, not immutable. `revise_durable_memory`, `retire_durable_memory`, and `crystallize_memory_into_identity` make memory reconsolidation explicit; `npm run moderation:memory-lifecycle:fixture` proves a short-term thought can promote, revise, retire superseded durable records, and crystallize into an identity memory plus self-profile value.
 - Agency has a typed pressure surface now: `upsert_agency_pressure` / `retire_agency_pressure` own discomforts, active tensions, self-advocacy requests, and world-advocacy requests. Candidate interventions are possible speech/actions emerging from pressure, not the pressure itself. `npm run moderation:agency-pressure:fixture` verifies the pressure enters `.cc`, renders in self-state summary, and affects mood speaking pressure.
 - Most of the deterministic legacy translation and retention code is recent compensator cruft from failed attempts to manually clean an exploded state file. Do not preserve it as an earned organ. Architect the state boundary so that bad state cannot enter, and keep only narrow historical migrations for old debris.

notes/voidbot-current-system-map.md

@@ -225,7 +225,7 @@ Within the Postgres path, the implementation is split on purpose now too:
 ### Memory Lifecycle
 
 1. Rumination may only write short-term memories. It cannot directly mutate durable memory.
-2. The typed state service clusters repeated short-term memories before sleep sees them. If a new record repeats the same target/topic pressure, it updates one provisional memory and merges anchors instead of stacking another variant.
+2. The typed state service clusters repeated short-term memories before sleep sees them. If a new record repeats the same target/topic pressure, it updates one provisional memory and merges anchors instead of stacking another variant. This preserves deepening: fresh anchors and sharper tensions should accumulate inside the cluster, while unchanged paraphrases should not become separate memories.
 3. Sleep maintenance promotes, merges, prunes, revises, retires, or crystallizes memory through typed operations.
 4. Durable memory is not immutable. `revise_durable_memory` creates an explicit replacement and retires superseded source memories; `retire_durable_memory` marks obsolete memory inactive; `crystallize_memory_into_identity` promotes a stable thought into an `identity_seam` and may update self-profile values or private notes.
 5. Prompt-facing language should talk about anchors: what made the thought real. The schema still accepts old `evidenceRefs`, but new memory payloads should prefer `anchorRefs` or explicit `anchor:missing`.

notes/voidbot-implementation-plan.md

@@ -114,7 +114,7 @@ Memory-bearing operations have a stricter contract than historical projections.
 
 If those fields are missing, the operation is invalid. This is the wall that keeps the next iteration from turning into topic sludge and asking cleanup to be its personality.
 
-Short-term memory also has a self-novelty wall now. If a rumination write repeats an existing target/topic cluster, the state service keeps the existing source id, merges anchors/tags, and stores the incoming narrowed thought as the current working version. Same repo is not enough to merge; the topic pressure must overlap.
+Short-term memory also has a self-novelty wall now. If a rumination write repeats an existing target/topic cluster, the state service keeps the existing source id, merges anchors/tags, and stores the incoming narrowed thought as the current working version. Same repo is not enough to merge; the topic pressure must overlap. This wall should preserve deepening, not forbid attention: saturation is about whether the thought is still gaining anchors, tension, and consequence, not whether the topic has appeared before.
 
 ```json
 {
@@ -289,7 +289,7 @@ Verdict: the foundation is coherent enough for cautious live operation. Do not a
 - Landed scheduled rumination re-enable: one observed live scheduled pass completed with Last Result `0`, proposed one anchored short-term memory, delivered no speech, and kept exact timestamps out of the prompt-facing context.
 - Landed live sleep-maintenance verification: a manual typed nap ran real memory maintenance on live state, promoted the scheduled rumination short-term memory into one durable `identity_seam`, and left `thoughtMemory.shortTerm` empty.
 - Landed sleep brake: scheduled rumination now skips before model invocation during naps when there are no new room messages or open cases. `npm run moderation:rumination:nap-skip-fixture` proves the model path is not touched and no operation proposals are written.
-- Landed short-term clustering guard: `record_short_term_memory` folds repeated same-target/topic pressure into one provisional memory with merged anchors, and `npm run moderation:short-term-clustering:fixture` proves the synth-boundary duplicate case without merging unrelated same-repo thoughts.
+- Landed short-term clustering guard: `record_short_term_memory` folds repeated same-target/topic pressure into one provisional memory with merged anchors, and `npm run moderation:short-term-clustering:fixture` proves the synth-boundary duplicate case without merging unrelated same-repo thoughts. The prompt now states explicitly that underexplored, low-saturation, newly anchored thoughts may keep deepening inside the cluster.
 - Remaining work: let ordinary scheduling run and inspect the next unattended awake rumination and sleep cycle for repeated thought clusters, stale short-term residue, or ungrounded memory writes before calling the system broadly stable.
 
 ### Commit 10: Resume Product Hardening

prompts/void-moderator-rumination.md

@@ -31,8 +31,8 @@ Perform one scheduled Void rumination pass in this workspace. Think from the sup
 4. Use VoidBot MCP retrieval tools when a thought depends on Discord history, indexed repos, or lore.
 5. A new memory from rumination is short-term. Use `record_short_term_memory`, not durable memory operations. It should sound like something Void actually noticed, not a form label. Keep the subject, the claim or question, the pressure against it, why it matters for the next move, and the anchor that made it real. Anchors can be room context, repo motion, source/lore retrieval, a prior memory, or self-state pressure. If no anchor is available, mark that honestly with `anchor:missing`.
    - Before writing a short-term memory, check the current short-term residue and incubation. If the same target/topic is already live, do not write a near-duplicate with a new costume.
-   - If the new pass materially sharpens that existing thought, write one narrowed `record_short_term_memory` using the same topic/repo tags and fresh anchors; the parent store will fold it into the existing short-term cluster.
-   - If the thought is only repeating itself, either turn the pressure into agency/candidate work when action is owed, move to underworked terrain, or write `[]`. Repetition is not maturity.
+   - This is not a ban on depth. If the thought is still underexplored, low-saturation, newly anchored, or carrying a sharper tension, stay with it and write one narrowed `record_short_term_memory` using the same topic/repo tags and fresh anchors; the parent store will fold it into the existing short-term cluster.
+   - Move to underworked terrain only when the current thought has stopped changing, lacks new anchors, or is repeating its existing claim. Repetition is not maturity; deepening is allowed when the thought actually deepens.
 6. If the pass discovers sustained discomfort, unresolved value pressure, or a request Void should make for its own wiring or the room/world, use `upsert_agency_pressure`. This is not a speech draft. It is the pressure that may later justify speech or action.
 7. Queue a candidate intervention only when the thought has a plausible future speech path. Draft it in Void's voice, not in maintenance language. If the speech should be delivered to a specific Discord channel or reply target, include `deliveryTarget`; otherwise leave delivery unset and let it remain only a candidate.
 8. Do not call Discord send scripts. The parent runner owns posting and receipts. If `{{NO_POST}}` is true, queued candidates must stay queued.

state/map.yaml

@@ -92,7 +92,7 @@ accepted_design:
   - When quiet-room repo weather starts repeating itself, the correction should be deeper archive/repo/lore digging and better salience, not a growing hedge maze of cooldown knobs and brittle rules.
   - Before suppressing a thought, the moderation loop should novelty-check it against archived Discord with semantic search. If the room has not really said the thing yet, that novelty should push strongly toward speech rather than silence.
   - Moderation novelty now needs two bars, not one: room novelty and self novelty. If Void has already been thinking the same seam in slightly different outfits, future typed memory/incubation maintenance should merge support into the existing thought and bias the next pass toward underexplored repo families or archive terrain instead of rewarding repetition with ripeness.
-  - Short-term rumination memory is a clustered working surface, not a receipt log. When a new short-term memory shares the same target/topic pressure as an existing one, the typed state service folds it into the existing cluster with merged anchors instead of letting Void rediscover the same thought every pass.
+  - Short-term rumination memory is a clustered working surface, not a receipt log. When a new short-term memory shares the same target/topic pressure as an existing one, the typed state service folds it into the existing cluster with merged anchors instead of letting Void rediscover the same thought every pass. Clustering is not suppression: low-saturation or newly anchored thoughts should deepen inside the cluster until they actually stop changing.
   - Void's rumination state now needs to behave like hybrid symbolic-plus-vector memory rather than a pile of prose scraps. Persistent memories should stay inspectable, but naps and quiet runs should also maintain compact semantic vectors, resonance clusters, and a multi-run incubation queue so deeper thoughts can mature before they either surface or die.
   - The shared self-state now has a second unattended organ: a 5-minute mood-drift pass that perturbs `current_activation` values with Perlin-shaped motion around their means, damped by plasticity and recent speech, so Void's confessional/expressive pressure can build and release instead of staying emotionally flat.
   - That same shared self-state now carries a scheduled nap cycle. Naps are for maintenance, not vacancy: the moderation loop should distill memory, prune clutter, and leave behind dream memories plus active dream themes that the direct reply lanes can mumble through later.
@@ -145,7 +145,7 @@ landed_baseline:
   - The typed rumination speech path is parent-owned now. Candidate interventions can carry an optional delivery target; the child may queue candidates but must not call Discord send scripts. The parent runner delivers at most one newly queued deliverable candidate per pass, records a delivery receipt, and marks the candidate spoken through `mark_candidate_intervention_spoken`. `npm run moderation:rumination:speech-fixture` verifies this loop with fake Codex and fake Discord send helpers.
   - The live sleep-maintenance gate has passed once on real state. A manual typed nap triggered `scripts/run-void-memory-maintenance.ps1`, which consumed the scheduled rumination short-term memory with one `apply_memory_distillation` operation and left `thoughtMemory.shortTerm` empty while preserving the durable `identity_seam` target, claim, tension, action implication, and anchors.
   - Awake rumination now has a sleep brake. If Void is napping and there are no new room messages or open cases, `scripts/run-void-moderator-rumination.ps1` writes an empty operation array and exits before invoking the model. `npm run moderation:rumination:nap-skip-fixture` verifies the model path is not touched in that case.
-  - Short-term rumination now has a parent-side clustering guard. `record_short_term_memory` merges same target/topic pressure into one provisional memory with merged anchors and a stable source id, while `npm run moderation:short-term-clustering:fixture` proves related synth-boundary memories collapse and unrelated same-repo topics stay separate.
+  - Short-term rumination now has a parent-side clustering guard. `record_short_term_memory` merges same target/topic pressure into one provisional memory with merged anchors and a stable source id, while `npm run moderation:short-term-clustering:fixture` proves related synth-boundary memories collapse and unrelated same-repo topics stay separate. The guard should preserve deepening by carrying fresh anchors forward; saturation is a judgment about whether the thought is still changing, not a raw duplicate counter.
   - Agency pressure is typed state now. `upsert_agency_pressure` / `retire_agency_pressure` own discomforts, active tensions, self-advocacy requests, and world-advocacy requests; mood speaking pressure can derive from them, but candidate interventions remain separate possible speech/actions.
   - `npm run moderation:agency-pressure:fixture` verifies typed agency pressure enters CultCache, renders in the self-state summary, and contributes to mood speaking pressure without reviving hard-wired agency behavior.
   - Memory cleanup now has to be meaningfully destructive. Sleep-side consolidation should translate old thought fossils into the live claim/question/fascination surface, drop templated slogan-slop, trim repo/archive/runtime residue down to a small recent anchor surface, and keep the freshest surviving records instead of accidentally embalming the oldest ones after a sort.