MECH-188 vs MECH-295 dual-path z_goal: read-site disambiguation

Created: 2026-05-08 Trigger: EXQ-536b found inject_observed_fraction=1.0 AND approach_commit_rate=0.0 co-existing across 3 seeds. Initial reading “downstream commit chain inert even with active z_goal” was partially correct but conflated two separate read sites.

The two paths

There are two write sites for z_goal-derived signals in select_action(), with different read sites and different gating predicates.

Path	Writer	Stored on	Read by	Gate predicate
MECH-188 inject (action-time)	`agent.select_action()` line ~2316–2321	`_goal_state_for_select` (a `GoalState.with_injection()` wrapper, distinct object from `self.goal_state`)	`e3.select()` per-candidate scoring via `cand_world_summaries` proximity	`cfg.goal.z_goal_inject > 0.0`
MECH-295 weak bridge (commit-driving)	`agent._e3_tick()` line ~3608 (`compute_anticipatory_liking_write`) and `select_action()` line ~2594–2642 (`compute_approach_cue_score_bias`)	reads `self.goal_state` (the persistent attractor)	composes `m295_bias` into `dacc_score_bias` (the negative score_bias E3 uses to favour approach)	`self.goal_state.is_active()` AND `goal_norm >= mech295_min_z_goal_norm_to_fire (default 0.05)`

GoalState.with_injection() (in ree_core/goal.py) creates a lightweight wrapper — a new GoalState instance whose _z_goal has a norm floor, sharing config and device with the original. It does not mutate self.goal_state._z_goal. The persistent attractor stays exactly where update_z_goal() left it.

GoalState.is_active() is self._z_goal.abs().sum() > 1e-6 — it checks the persistent attractor, not the wrapper.

Why EXQ-536b reads as “inert”

EXQ-536a established that update_z_goal() never seeds: the SD-012 drive multiplier collapses to ~1.0 at the moment of contact (energy=0.995 → drive=0.005), so effective_benefit = 0.036 < benefit_threshold (0.1). Persistent attractor stays zero.
self.goal_state._z_goal stays at the zero vector for the whole run.
self.goal_state.is_active() returns False for the whole run.
The MECH-295 bridge (both write site at _e3_tick and read site at select_action) is gated by is_active(). Both bridge call sites short-circuit.
Meanwhile, cfg.goal.z_goal_inject = 0.3 ARM_1 builds _goal_state_for_select via with_injection(0.3). This wrapper has _z_goal.norm() >= 0.3. E3.select sees it in per-candidate proximity calc → inject_observed_fraction = 1.0.
But the m295 bias is 0.0 (bridge skipped), so dacc_score_bias carries no approach pressure. approach_commit_rate = 0.0.

The “inert downstream chain” reading was therefore wrong-by-conflation. The chain is not broken; it is bypassed. The bypass is intentional from MECH-188’s design intent (the docstring says “applies to action selection only, does NOT modify the persistent attractor”) but unintentional as a force-arm test of MECH-295.

What this means for the diagnostic queue

EXQ-536b cannot falsify the MECH-295 weak-bridge reading. It tested the wrong path. Three follow-ons would discriminate cleanly:

EXQ-536c (MECH-295 direct force-arm): seed self.goal_state._z_goal directly to norm 0.3 at episode start, bypassing update_z_goal(). If approach_commit_rate lifts, the bridge is intact and the upstream seeding (SD-012 drive-vs-benefit) is the sole blocker. If it stays at 0, MECH-295 is also broken.
EXQ-536d (lower min_z_goal_norm_to_fire): drop mech295_min_z_goal_norm_to_fire from default 0.05 to 0.01. Tests whether the bridge’s internal threshold is too high for naturalistic z_goal seeding when contacts are sparse.
EXQ-536e (combined): both at once. Cleanest single-shot test.

A fourth option is structural: extend with_injection() to also write the floor into the persistent attractor, or have MECH-188 set both _goal_state_for_select and self.goal_state._z_goal. This would make the MECH-188 inject a true upstream force-arm. Worth considering as a small SD amendment if EXQ-536c/d/e show MECH-295 is intact.

Implication for the EXQ-483 catatonic-lock diagnosis

The MECH295LikingBridge constructor docstring at agent.py:628-637 cites EXQ-483 as the empirical anchor for “drive amplification produces a passive z_goal latent without behavioural consequence.” That diagnosis is unchanged — but the EXQ-483 cohort scripts also did not seed the persistent attractor in some arms (they relied on the same benefit-threshold path that EXQ-536a just falsified for sparse contact regimes). Whether EXQ-483’s “approach_commit = 0.0 across all arms” was due to bridge inertness or due to upstream seeding failure is not yet disambiguated by the existing evidence. The EXQ-536c/d/e cohort would clear this up retroactively.

Companion claim updates needed

When EXQ-536c/d/e land, add a note to MECH-295’s evidence_quality_note clarifying that EXQ-536b is non_contributory for MECH-295 directly (it tested MECH-188 path only). Currently the manifest is correctly tagged non_contributory for diagnostic purposes, but the implication for MECH-295 should be made explicit so the V3-pending note for MECH-295 doesn’t silently absorb 536b as “weakens.”