ASSERT_MY_TAPROOT removed: one identity assert

[EvanWinget]

The follow-up PR 62 scheduled: ASSERT_MY_TAPROOT (0x37) leaves the vocabulary, ratified as condition-record decision 29 (decision by Evan, 2026-08-23, at the PR 62 review, the venue decision 28 reserved for this call). The objection register also gains the hardware-signer note from the same review conversation.

Why removal loses nothing. ASSERT_MY_TAPTREE proves the same fact at 200 instead of 1,300,200, and the two capabilities that leave with 0x37 are unreachable: the empty-root form asserts a treeless output on a spend that by definition executes a leaf, and exploiting its tweak-collision tolerance requires a discrete log. PR 62 removed the last caller, so the entry would have carried review burden with no user. CREATE_OUTPUT_TAPROOT and the derivation machinery are untouched.

Read the commits in this order

1. spec: ASSERT_MY_TAPROOT leaves the vocabulary. The table row and entry go, the self assert family is five conditions, 0x37 joins the 0x34 to 0x36 gap (invalid, not reserved), and the taptree entry restates its agreement fact against ASSERT_MY_SCRIPTPUBKEY over the derived scriptPubKey.
2. conditions: drop the removed assert's implementation. Constant, dataclass, parse function, dispatch branch, cost row, validation branch, display and name tables. The language has 26 condition constants.
3. vectors: 0x37 re-pins as a gap. 26 cases leave, gap_0x37_invalid arrives, the all-opcodes total re-pins at 14,405,200 (exactly 1,300,200 lower), and the covenant recreation-missing case re-pins on the taptree assert with an honest identity.
4. tests: the taptree assert's oracle is the tweak derivation directly. The one substantive test change: the agreement property now pairs ASSERT_MY_TAPTREE with ASSERT_MY_SCRIPTPUBKEY over the scriptPubKey its operands derive, keeping the BIP341 derivation as the identity assert's oracle with no intermediate assert.
5. docs: decision 29 recorded. Record, C24, comparison table, glossary, plan bullet.
6. docs: the register anticipates the hardware-signer objection. In scope by request: this PR was asked to carry the removal and this note together. NUMS-keyed spends sign outpoint-bound condition digests, not a BIP341 sighash, and shipping hardware signers refuse arbitrary digests. An xpub-only watcher assembles everything else. Phase 4 owes the integration note.

What a reviewer should scrutinize

• The oracle chain survives the removal. The condition record's rule 2 column had the taproot assert as the taptree assert's oracle through the agreement invariant. That chain is now direct: the shared BIP341 derivation anchors the taptree assert through the derived-scriptPubKey agreement invariant, stated in VALIDATION.md's invariants, the record, and pinned by the rewritten hypothesis property.
• The gap is visible. gap_0x37_invalid expects bad_condition_opcode, the same shape as the 0x39 gap vector, so a future revival of 0x37 is a loud re-pin, never an accident.
• The all-opcodes cost total. Old 15,705,400, new 14,405,200. The conditions payload changed only by deleting the one 0x37 element, checked by decoding both.

Verify independently

.venv/bin/pytest python/tests                # 812 passed
.venv/bin/python tools/run_vectors.py        # 38 files, 1106 cases, 0 failures
ci/lint/lint.sh                              # all checks passed
grep -rn "ASSERT_MY_TAPROOT" spec/ python/ tools/ vectors/   # no hits
grep -rn "ASSERT_MY_TAPROOT" docs/           # historical mentions only: the record, the comparison row, the puzzle docs' saving sentences, the plan

Confidence

High, 9 of 10. The removal surface was enumerated by grep before editing and re-checked after, every gate is green, and the one deliberate semantic artifact (the gap vector) is pinned. The vendored Bitcoin Core oracle files were touched by a formatter mid-session and restored to main's blobs before any commit, git diff main -- tools/oracle is empty. Residual risk sits in wording drift across the spec and record, which review reads catch best.

[EvanWinget]

Review pass (code-review skill, high, Fable finders, adversarial verify)

Eight finder angles, nine deduplicated candidates, one verifier each. All nine confirmed at least in substance and all are folded in the review fold-in commit:

1. COSTS.md still said "the two derivation entries" in the section 10 charge-order sentence. Singular now.
2. The condition record's rule 5 row still claimed vectors for both derivation entries' width and derivation defects after the assert-side twins were deleted. Restated for the one entry.
3. The puzzle docs' saving baselines compared against ASSERT_MY_TAPROOT in present tense. Both now name it as the derivation assert decision 29 has since removed.
4. gap_0x37_invalid sat where the deleted parse cases were, away from the contiguous gap block. Moved: the block now reads 0x34, 0x35, 0x36, 0x37, 0x39.
5. HONEST_SPKS re-derived tweaks the module already held as SPK constants, contradicting the derived-once comment. The dict is now the single derivation site and the constants read from it.
6. The cost-invariants menu comment undercounted its generic entries after the taptree entry joined. Reworded.
7. GOOD_INTERNAL_KEY was dead once its only test left. Deleted.
8. A vestigial f-string prefix survived in the validation clause after its placeholder left. Plain literal now.
9. The register's owed-by-phase list missed the new hardware-signer note. Phase 4 now carries it beside the section 5 bullet. The scope half of this finding is answered in the PR body: the note was requested together with the removal.

Final: 812 tests, 1106 vector cases, lint clean, CI green.