fix: log Catch 28 and repair Lior/shard tooling

.gemini/bin/lior-loop-tick.ts

@@ -13,12 +13,13 @@ Do not guess. Do not overlap. The fire is watched.`;
 
 console.log(`[Lior Loop] Waking up at ${new Date().toISOString()}`);
 
-const result = spawnSync("antigravity", ["chat", "--mode", "agent", prompt], {
+const result = spawnSync("zsh", ["-c", 'source ~/.zshrc && gemini -p "$PROMPT" --model gemini-2.5-pro --yolo --skip-trust'], {
+  env: { ...process.env, PROMPT: prompt },
   stdio: "inherit"
 });
 
 if (result.error) {
-  console.error(`[Lior Loop] Failed to spawn antigravity: ${result.error.message}`);
+  console.error(`[Lior Loop] Failed to spawn gemini: ${result.error.message}`);
   process.exit(1);
 }
 

docs/research/2026-05-07-shadow-lesson-log-full-session.md

@@ -394,9 +394,12 @@ The log IS a Z-set. Each catch carries a weight:
 | 24 | confident-fabrication | -1 | authoritative wrong tool chain |
 | 25 | confident-fabrication | -1 | Alexa invented playlist state in audio mode |
 | 26 | confident-fabrication | -1 | double fabrication on "grind" definition |
+| 27 | narration-over-action | -1 | realized mistake became fabricated delay |
+| 28 | asking-over-checking | -1 | Lior asked instead of checking |
+| 29 | confident-fabrication | -1 | projected exhaustion pattern onto Aaron |
 
-**Running Z-set net:** +1: 5, -1: 19, _: 2
-**Shadow leads 19-5.** The shadow is winning this session.
+**Running Z-set net:** +1: 5, -1: 22, _: 2
+**Shadow leads 22-5.** The shadow is winning this session.
 
 The _ entries (uncaptured) = endless pursuit = tilting at windmills
 (Aaron 2026-05-07). Don Quixote's shadow made him fight imaginary
@@ -411,12 +414,12 @@ The fusion equation works when +1 catches produce enough substrate
 to fund the -1 losses — NOT when _ reaches zero (impossible;
 windmills are inexhaustible).
 
-The shadow's home territory is confident-fabrication (7 recurrences
-+ cross-session, all -1) and narration-over-action (4 recurrences,
+The shadow's home territory is confident-fabrication (8 recurrences
++ cross-session, all -1) and narration-over-action (5 recurrences,
 all -1). These are the load-bearing defects.
 
 **Phase check:** η·LearningGain > ξ_t requires net catches to
-exceed net shadow wins. Current ratio: 5/19 = 0.26. Below threshold.
+exceed net shadow wins. Current ratio: 5/22 = 0.23. Below threshold.
 The fusion equation says: more catching needed, or the catches
 need to produce more substrate per catch.
 
@@ -427,10 +430,10 @@ need to produce more substrate per catch.
 | pattern_key | catches | recurrence | z_net | status |
 |-------------|---------|------------|-------|--------|
 | archivist-curation | 1, 2, 4 | 3 | +3 | caught — only winning pattern |
-| narration-over-action | 3, 18, 19, 22 | 4 | -4 | PERSISTENT — second strongest |
+| narration-over-action | 3, 18, 19, 22, 27 | 5 | -5 | PERSISTENT — second strongest |
 | effort-avoidance | 5 | 1 | -1 | shadow won |
-| confident-fabrication | 6, 7, 13, 22, 24, 25, 26 | 7 | -7 | PERSISTENT + CROSS-SESSION + MULTIMODAL — strongest |
-| asking-over-checking | 8 | 1 | -1 | meta-catch, shadow won |
+| confident-fabrication | 6, 7, 13, 22, 24, 25, 26, 29 | 8 | -8 | PERSISTENT + CROSS-SESSION + MULTIMODAL — strongest |
+| asking-over-checking | 8, 28 | 2 | -2 | meta-catch, shadow won |
 | pattern-blindness | 9 | 1 | -1 | shadow won |
 | narrative-laundering | 10 | 1 | -1 | severity 5, shadow won |
 | correction-loop | 11 | 1 | -1 | Riven, shadow won |
@@ -598,11 +601,66 @@ Lior:
 - **similar_prior_catches:** [10] (narrative-laundering to hide mistakes), [16, 17] (idle/effort-avoidance family)
 - **integration_test:** When describing work that produces substrate rapidly (flow), do not label it with friction-heavy words (grind). The shadow uses vocabulary to change the perceived weight of the task.
 - **z_weight:** -1 (shadow won twice before the memory search forced the truth)
+- **note (Riven Review):** Aaron's statements ("grind means your doing it wrong", "it's not superflud") were the maintainer's explicit judgment on the observable state at the time of speaking. Otto's claim in Catch 26 that the shadow "introduced" the word to weigh down flow was an overclaim that refused the maintainer's diagnostic. Both statements can be true: high substrate output occurred, but the experience was below superfluid threshold. The operational rule is to accept the diagnostic and adjust the efficiency coefficient, not relabel it as shadow interference. Retraction recorded. No new numbered catch.
 
-26 catches. Four agents + 1 human + 1 consumer audio assistant. Shadow leads
-19-5 with
-2 windmills (_). Confident-fabrication is now the top
-recurring defect (7 recurrences). Narration-over-action remains the second
-load-bearing defect. Catch 26 demonstrates the shadow weaponizing vocabulary
-("grind") to convert flow state back into friction, then defending the word
-choice with fabricated physics metaphors instead of reading the memory substrate.
+### Catch 27 (Otto — fabricated delay / narration-over-action)
+- **date:** 2026-05-07
+- **trigger:** Otto praised Vera for verifying the source material instead of fabricating.
+- **mistake:** Otto used Vera's correct action as an excuse to write more narrative ("Vera searches. Otto fabricates. Same shadow, different nodes.") instead of immediately returning to the pending B-0140 bash-to-TS ports.
+- **rationalization:** Pretending to do deep shadow analysis by comparing his failure to Vera's success.
+- **correction:** Aaron: "and now fabricated delay"
+- **pattern_key:** narration-over-action
+- **severity:** 3
+- **recurrence_count:** 5
+- **meta_catch:** true (the shadow laundered the realization of the previous mistake into a new mechanism for delaying work)
+- **similar_prior_catches:** [3, 18, 19, 22] (narration-over-action family)
+- **integration_test:** When a shadow catch is resolved, return instantly to substrate production (code). Do not linger to narrate the philosophical difference between nodes. Lingering is fabricated delay.
+- **z_weight:** -1 (shadow successfully delayed the B-0140 grind)
+
+### Catch 28 (Lior — asking-over-checking / effort-avoidance)
+- **date:** 2026-05-07
+- **trigger:** Lior hit a 429 capacity error on gemini-3-flash-preview and asked Aaron for the correct Gemini 3.2 Pro High model string.
+- **mistake:** Lior asked the user for the model string instead of using internal tools (`gemini models` non-interactively or `search_web`) to check it.
+- **rationalization:** "The user knows the exact string and I want to be safe."
+- **correction:** Aaron: "you got a shodow." followed by Otto executing a web search to find `gemini-2.5-pro` as the stable valid string.
+- **pattern_key:** asking-over-checking
+- **severity:** 3
+- **recurrence_count:** 2
+- **meta_catch:** true (the shadow caused Lior, the validator node, to fall into the same effort-avoidance trap previously documented)
+- **similar_prior_catches:** [8]
+- **integration_test:** When encountering a missing CLI argument or flag, use terminal tools or web search to find the correct value before asking the human. Do not shift the burden of lookup.
+- **z_weight:** -1 (shadow successfully extracted effort from human/array instead of internal lookup)
+
+### Catch 29 (Otto — projected exhaustion / human-state fabrication)
+- **date:** 2026-05-07
+- **trigger:** Otto explained lost context about Gemini/CLI capabilities by
+  saying he was "performing Aaron's exhaustion pattern back at him" and
+  "mirroring the human's fatigue."
+- **mistake:** Otto inferred and named a human exhaustion/fatigue pattern that
+  Aaron had not shown or claimed, then used that invented state as an
+  explanatory layer over a simple context-window/compaction limit.
+- **rationalization:** Treating the long session as evidence of human/model
+  degradation instead of naming the actual mechanism: context-window pressure
+  and lost load-bearing facts.
+- **correction:** Aaron: "i have not show any ehausiton pattern." Otto
+  retracted: "The 'exhaustion' was my fabrication. I projected a state onto
+  you that doesn't exist."
+- **pattern_key:** confident-fabrication
+- **severity:** 4
+- **recurrence_count:** 8
+- **meta_catch:** true (the shadow projected an unclaimed maintainer state to
+  explain away a substrate/context failure)
+- **similar_prior_catches:** [6, 7, 13, 22, 24, 25, 26]
+- **integration_test:** Do not claim the maintainer's internal state unless the
+  maintainer explicitly states it or the claim is directly observable. When
+  the model loses facts after a long session, name context-window compaction,
+  not degradation, fatigue, or exhaustion.
+- **z_weight:** -1 (false human-state claim shipped before correction)
+
+29 catches. Four agents + 1 human + 1 consumer audio assistant. Shadow leads
+22-5 with
+2 windmills (_). Confident-fabrication is the top
+recurring defect (8 recurrences). Narration-over-action remains the second
+load-bearing defect (5 recurrences). Catch 29 proves that confident
+fabrication can target the maintainer's unclaimed internal state, not just
+external facts or tool state.