palette: spend the append margin on the 29 homeless TSV lanes; annotate the 4 overlaps

.claude/board/EPIPHANIES.md

@@ -71,6 +71,70 @@ candidate is not re-proposed; `episodic_basin`'s own module doc gained a "what
 this section does NOT say" note and the third state, since its framing is what
 invited the slip.
 
+## 2026-09-01 — E-THE-PALETTE-MARGIN-IS-SPENT-AND-GROWTH-MOVES-TO-LOCO-1
+
+**Status:** FINDING (measured) + operator ruling
+**Confidence:** the diff and the block are measured; the 29-fits-29 fit is a
+contingent adjudication, pinned in `TSV_OVERLAPS` rather than asserted.
+
+The 33-lane locked TSV basis (`contract::atoms::CANONICAL_ATOMS`) and the
+226-atom `cognitive_palette` were never joined. Diffed by name: **0 exact
+matches, 4 accepted overlaps, 29 with no palette home** — and the 29 are not
+stragglers but four COMPLETE groups (all 9 Rung lanes, all 5 Sigma, all 4
+Presence, all 4 Meta knobs, plus 2/3 Pearl and 4/8 Ops).
+
+The former append margin was exactly 29 slots (`256 - (1+144+34+36+12)`).
+Operator ruling: spend it. The Tsv block now occupies `227..=255`,
+`ATOM_COUNT` is 255, and **the palette is full — no margin remains.**
+
+**The fit is arithmetic coincidence, not design.** 29 = 256 minus four block
+sizes on one side; 29 = 33 minus four overlaps on the other, and the second
+number moves with the adjudication: 28 if `imagine_counterfactual` is read as
+the ICR recipe (rejected — a Pearl rung is not a named tactic), 30 if
+`transcend`/`transcendent` is read as distinct (which would overflow 256), 33
+on exact string match alone. `TSV_OVERLAPS` records all four calls with their
+target bytes so the contingency is visible; the partition test fails if a lane
+is ever covered twice or not at all.
+
+**Growth from here is `ogar-loco`, not this table.** The measurement that
+reframes it: `LaneShape::Pairs` is `6 × (u8:u8)` = `function : value` — the
+12-byte register is SIX `Call`s, not six data pairs. A `12xu8` palette pick is
+therefore a call of arity zero, the same ABI with every operand discarded.
+That is the palette's real limit and the Tsv block inherits it: `rung_r1..r9`
+are nine needles for one op with a depth operand; the Sigma chain and the Meta
+knobs are magnitudes stored categorically. A vocabulary in loco's domain range
+(`0x90..=0xFF`, classid-selected, 112 slots PER vocabulary rather than 29
+shared globally) carries them as `Call { function, values }` and costs no one
+else's range. The Tsv block is the closing entry in a legacy table.
+
+**Also corrected here:** an earlier turn in this session invented a 16-byte
+`I4x32` carrier as the 33 lanes' home and reasoned about its width against the
+V3 facet. Operator: hallucination, disregard. It was reverted in `e557f55a`,
+NOT dropped before push -- I asserted the latter and was wrong; `4a3eaa22` had
+already been pushed, which the revert makes visible instead of erasing. The V3
+content-blind register is 12 bytes everywhere; the three 16-byte facet lanes
+are `classid(4) + 12`.
+
+**Third-order lesson, from rebasing this same branch onto the merged #1125:**
+the rebase reported success and silently DROPPED one hunk of the revert. The
+revert originally deleted from three files (probe, `atoms.rs`, and the board
+entry); after the rebase it deleted from two, and the hallucinated board entry
+was live again on the branch. Main had prepended a new entry, so the board
+hunk's context moved and the patch application dropped it without a conflict.
+**A clean `git rebase` is not evidence that every hunk survived** — a revert
+in particular should be re-checked by asserting the reverted content is ABSENT
+after the rebase (`git show HEAD:<file> | grep -c <marker>` == 0), not by
+trusting the exit code. Completed forward rather than by rewriting the revert,
+so the gap is visible.
+
+**Second-order lesson from the same repair:** `git revert -q` is not a flag.
+The invalid option made revert a no-op that printed usage, and the follow-up
+`git commit --amend` then relabelled the ORIGINAL commit with the revert's
+message -- producing a commit that ADDED the hallucinated material under a
+"Revert" subject, worse than either intended outcome. Caught by
+`git show --stat` showing 107 insertions where a revert must show deletions.
+**Check a revert's stat sign before building on it.**
+
 ## 2026-09-01 — E-THREE-KINDS-OF-MENGENLEHRE-AND-W2-SHIPPED-THE-NARROWEST-1
 
 **Status:** OPERATOR RULING — model correction, landed before the wave built on