probe(helix): the ruler reads the address as a NUMBER — intake blindness kills the residue sweep, and "stride 4 isn't golden" is refuted

[AdaWorldAPI]

PROBE-CLAM-VS-HELIX-RESIDUE leg 0, following the merged #875 (E-WORDNET-MAKES-THE-4-ARY-ADDRESS-SEMANTIC-1) and the operator's division of labour: the 4⁴ fold is the ADDRESS, CLAM is the established CALCULATOR, HHTL + helix residue is the alternative — "exact address, exact spacial location / perturbation".

The queued sweep did NOT run, and not running it is the result

The hypothesis was: a finer address makes more of the answer deterministic PLACE and less of it stored RESIDUE. Two failure modes were pre-registered before any measurement. Running them first was decisive.

H1 — the carving cannot reach the ruler (pre-registered (a): CONFIRMED, fatal)

• Type-level (stated, not measured): from_hhtl(path: u64, depth: u8) has no parameter that could carry a carving, so level structure has no channel into from_place(p) = p % 17.
• Empirical (the gate): the entire difference between the two carvings is a constant rotation of 2 across all 256 cells, variance 0. A cell-dependent shift is the falsifier.

A residue-vs-granularity sweep therefore cannot carry signal; its null would say nothing about 4⁴, and reporting "4⁴ doesn't help the calculator" from it would have been a false negative manufactured by the seam. The 4⁴ address result (24.71× out-of-cell band lift, 2.47-hop sub-nibble gap) is untouched.

H2 — the "stride 4 clumps" suspicion is NOT CONFIRMED (no design claim replaces it)

Gap spread (max−min; lower = more uniform):

n	stride 4 (shipped)	stride 11 (φ)
2	9	5
3	5	1
4	1 — gaps [4,4,4,5]	5
5	3	4
6	3	4

⚠ The design conclusion is WITHDRAWN. The premise that a 4-ary tier consumes four ruler steps is not wired into anything that ships: ResidueEncoder::encode reads only start_offset (n=1, residue.rs:157); index/arc appear solely in walk_spectrum and the crate's own tests; the shipped NiblePath is FAN_OUT = 16. The n=4 result is a property of a hypothetical consumer. This retracts a suspicion; it does not establish that the constants are "matched to 4-ary".

H3 — the blindness is FIXABLE (a direction, not a dead end)

Per-tier seeding (one ruler per 2-bit group, folded with its level). Exhaustive over all 256·255/2 = 32,640 cell pairs — no sampling, no thinning.

The acceptance criterion is a declared null model, not a threshold — H0 = "the flat seed carries no ancestry information", so its agreement rate among eligible pairs must not exceed its rate among all pairs:

	agreement rate
flat intake, eligible pairs (8,064)	360 / 8,064 = 4.4643 %
flat intake, all pairs (32,640) — the base rate	1,800 / 32,640 = 5.5147 %
Δ	−1.05 pp
per-tier intake, eligible pairs	8,064 / 8,064 = 100 % by construction

The flat seed gives no ancestry lift at all — and is in fact mildly ancestry-AVERSE, because cells sharing an address prefix are numerically close and therefore less likely to be congruent mod 17. Falsifier: an eligible rate materially above base, which would mean the flat seed does carry ancestry. The 100 % is asserted as a construction check, explicitly not as evidence.

The unblock is to feed the ruler the address's hierarchy, not its integer value — a helix API change needing its own review, not done here.

Eight defects across the arc, every one in a falsifier or a label — none in a measurement

1. H1 compared SORTED histograms — sorting hid that every cell had moved. (codex P1)
2. H2 drew a design conclusion from an unwired premise — withdrawn everywhere it appeared. (codex P1)
3. H3 counted before the eligibility check — "4,662 pairs" was every thinned pair; the "≈23×" was arithmetic on a mislabelled denominator. (codex P2)
4. The fix for Claude/review lance graph architecture i6 t kf #3 repeated the error one level up — 1,152 was the eligible thinned sample, not the population. Root cause of both: a thinning that bought nothing. Removed rather than relabelled. (CodeRabbit)
5. H3's threshold was hand-tuned — flat_rate < 0.25 would have passed at 24.9 % while the prose claimed ≈1/17. Replaced with the declared null above. (CodeRabbit)
6. …and that null immediately failed my first formulation ("the two rates coincide"), which was itself wrong — they diverge by −1.05 pp. The gate now asserts only the true, weaker claim. Self-caught by making the criterion explicit.
7. The EPIPHANIES title kept the withdrawn conclusion the body had already retracted. (CodeRabbit)
8. H3's first run sampled 0 pairs (256 as u8 == 0 made the loop empty) and correctly FAILED rather than passing on zero evidence. (self-caught)

The transferable rules

• A calculator's INTAKE can silently discard the very structure a probe is trying to measure. Check what the consumer actually reads before sweeping over what the producer emits.
• Fixing a mislabelled denominator can reproduce the mislabel one level up. When a sample keeps generating label errors and full enumeration is cheap, delete the sampling.
• A hand-tuned threshold cannot fail informatively; a declared null can. Twice here, stating the criterion explicitly exposed that the replacement claim was also wrong.

Core-First compliance

CONSUMES helix::curve_ruler::CurveRuler — nothing re-derived.

Boundaries

• Properties of the RULER only (integer arithmetic over 17). No residue encoded, no CLAM comparison, no cost measured.
• Brief-gates AGREEMENT and COST remain NOT RUN — blocked on H1, not skipped.
• H2 uses gap spread, a crude discrepancy proxy; it does not claim stride 4 is optimal.

Contents

• crates/helix/examples/probe_hhtl_intake_blindness.rs — the probe: 3 gates + a #[cfg(test)] module of 7 focused tests (including a falsifier that a structure-carrying seed would not be constant-shift, and a check that stride 4 over 16 covers only 4/16 — the comma, pinned as a test).
• Board: EPIPHANIES.md (E-THE-RULER-READS-THE-ADDRESS-AS-A-NUMBER-NOT-A-HIERARCHY-1) + exec-runs/hhtl-intake-blindness.md, both carrying dated ⊘ correction markers.

Additive only — no change to helix's shipped API.

Verification: 3/3 gates green; 7/7 unit tests; cargo clippy --manifest-path crates/helix/Cargo.toml --example probe_hhtl_intake_blindness -- -D warnings clean (example-scoped per repo practice — --all-targets pulls in pre-existing lint debt in helix's own lib tests, unrelated to this diff); cargo fmt applied, with unrelated fmt drift in two other helix examples restored rather than carried.

---

🤖 Generated with Claude Code

https://claude.ai/code/session_01K3RyLEbuNSHxxB3NTTrGki

[AdaWorldAPI]

@coderabbitai review

Requesting while still draft — merges have repeatedly outrun reviews in this repo, so findings are worth more now.

Three places where this probe is most likely to be wrong, since they're the load-bearing claims:

1. H1's equivalence argument. I claim the two carvings are indistinguishable to the ruler because from_place(p) = p % 17 sees only the integer. The histograms are a pure rotation, which is consistent with that — but a rotation is also what you'd get from any pair of depths, so please check whether the gate actually establishes structure-blindness or merely re-derives that (cell + depth) % 17 shifts with depth. If it's the latter, the conclusion may still hold but the gate is weaker than the claim.
2. H2's metric choice. Gap spread (max−min) is a crude discrepancy proxy; star-discrepancy or the full three-distance gap multiset might rank the strides differently at n=5,6 where the margin is only 1. The n=4 result (spread 1 vs 5) is the load-bearing one and looks robust, but I'd like the weaker rows challenged.
3. H3's thinning. Pairs are sampled by (a+b) % 7 == 0, which is not independent of address structure — a+b correlates with low-order bits, i.e. with the very 2-bit groups the per-tier intake keys on. If that biases the 23× ratio, the direction probably survives but the magnitude shouldn't be quoted.

Also worth flagging what this PR deliberately does not claim: no residue was encoded, no CLAM comparison ran, and no cost was measured — the AGREEMENT and COST gates are blocked on H1, not skipped.

---

Generated by Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 506514dc78

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[AdaWorldAPI]

All three findings confirmed and fixed in a096114. None was fixed by adjusting a threshold, and the second one means a claim in my own PR description was wrong.

P1 #1 — H1 sorted the histograms. Correct, and worse than it first looks: the unsorted histograms genuinely differ (peak at index 4 vs 2), so sorting made the gate pass while hiding that every cell had moved. An aggregate distribution cannot establish per-cell invariance. Replaced with the argument that actually holds, in two halves:

• type-level, stated not measured — from_hhtl(path: u64, depth: u8) has no parameter that could carry a carving, so level structure has no channel into p % 17. No measurement can establish this and the probe no longer pretends to;
• empirical, and this is the gate — the entire difference between the two carvings is a constant rotation of 2 across all 256 cells, variance 0. A cell-dependent shift is the falsifier.

P1 #2 — H2's design conclusion rested on an unwired premise. Withdrawn. I verified your call-site claim against the source and it holds exactly: ResidueEncoder::encode reads only start_offset (n=1, residue.rs:157); index/arc appear solely in walk_spectrum and the crate's own tests; the shipped NiblePath is FAN_OUT = 16. Nothing that ships consumes a 4-step prefix, so the n=4 comparison is a property of a hypothetical consumer, not evidence about the shipped design. "The shipped constants are MATCHED to the 4-ary use case" is retracted from the probe, the module docs, the exec-run, and the epiphany. What survives is narrower and still worth recording: the pre-registered suspicion is NOT CONFIRMED — a retraction of a concern, not a vindication of a design. That distinction is exactly what I'd blurred.

P2 #3 — H3 published a false sample size, and your inference about the true one was right. The counter incremented before the k > 0 check, so "4,662 pairs sharing ≥1 address level" was every thinned pair. Since tiered agreement holds for every eligible pair by construction, 1,152 is the eligible population. Corrected:

	flat intake	per-tier intake
ancestry preserved	51 / 1,152 = 4.4 %	1,152 / 1,152 = 100 %

4.4 % sits essentially at the 1/17 ≈ 5.9 % chance level — the honest reading is "the flat intake keeps ancestry about as often as chance would", which is a better statement than the ≈23× ratio I published. That ratio was arithmetic on a mislabelled denominator and is retracted. The 100 % is now asserted as a construction check, explicitly not as evidence, since it cannot be otherwise.

Both board records carry dated inline ⊘ correction markers per the append-only rule, and the PR description above still shows the pre-correction H3 numbers — I'll update it next so the most-read artifact isn't the stale one.

3/3 gates green, clippy -D warnings clean, fmt applied.

---

Generated by Claude Code

[coderabbitai]

📝 Walkthrough

Walkthrough

The pull request adds a Rust probe for HHTL intake blindness. It tests carving rotation, stride gap spread, and hierarchical seeding. It records execution results, corrected statistics, validation commands, scope limits, and follow-up API work.

Changes

HHTL intake probe

Layer / File(s)	Summary
Probe harness and measurements crates/helix/examples/probe_hhtl_intake_blindness.rs, .claude/board/exec-runs/hhtl-intake-blindness.md	The new executable defines H1, H2, and H3, reports gate results, builds start-offset histograms, and measures prefix gap spread.
Intake and seeding gates crates/helix/examples/probe_hhtl_intake_blindness.rs, .claude/board/exec-runs/hhtl-intake-blindness.md	The probe checks constant rotation, compares strides 4 and 11, and compares flat seeds with per-tier hierarchical seeds.
Verdict and execution records crates/helix/examples/probe_hhtl_intake_blindness.rs, .claude/board/exec-runs/hhtl-intake-blindness.md, .claude/board/EPIPHANIES.md	The probe skips the residue sweep, exits nonzero when gates fail, and records corrected measurements, verification results, scope limits, and follow-up actions.

Estimated code review effort: 3 (Moderate) | ~30 minutes

Possibly related PRs

• AdaWorldAPI/lance-graph#485: Related HHTL, CurveRuler, address-carving, and stride-4 behavior.

Poem

A rabbit checked the helix path,
Found flat seeds missed the branch.
Tier by tier, the roots align,
Three gates pass in ordered time.
The residue sweep waits its turn.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title accurately identifies the probe's main finding about intake blindness and the withdrawn stride-4 claim, although it is longer than preferred.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
set -euo pipefail

printf '%s\n' '== Changed files =='
git diff --name-only "$(git merge-base HEAD origin/main 2>/dev/null || git merge-base HEAD main 2>/dev/null || git rev-list --max-parents=0 HEAD | tail -n1)" HEAD 2>/dev/null || true

printf '%s\n' '== Candidate files =='
fd -t f -a 'probe_hhtl_intake_blindness\.rs|curve_ruler\.rs|residue\.rs|lib\.rs' crates/helix 2>/dev/null || true

printf '%s\n' '== Probe outline =='
ast-grep outline crates/helix/examples/probe_hhtl_intake_blindness.rs --items all 2>/dev/null || true

printf '%s\n' '== Curve-ruler outline =='
ast-grep outline crates/helix/src --items all --match 'CurveRuler|ResidueEncoder|NiblePath' 2>/dev/null || true

printf '%s\n' '== Relevant identifier uses =='
rg -n -C 3 '\b(from_hhtl|from_place|start_offset|walk_spectrum|\.index\s*\(|\.arc\s*\(|FAN_OUT)\b' crates/helix --glob '*.rs' 2>/dev/null || true

Length of output: 21701

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail

printf '%s\n' '== Probe: all gate code =='
nl -ba crates/helix/examples/probe_hhtl_intake_blindness.rs | sed -n '75,310p'

printf '%s\n' '== CurveRuler construction =='
nl -ba crates/helix/src/curve_ruler.rs | sed -n '18,68p'

printf '%s\n' '== ResidueEncoder call path =='
nl -ba crates/helix/src/residue.rs | sed -n '125,180p'

printf '%s\n' '== PR diff: probe =='
git diff "$(git merge-base HEAD origin/main 2>/dev/null || git merge-base HEAD main 2>/dev/null || git rev-list --max-parents=0 HEAD | tail -n1)" HEAD -- crates/helix/examples/probe_hhtl_intake_blindness.rs

printf '%s\n' '== Changed board claims =='
git diff --unified=12 "$(git merge-base HEAD origin/main 2>/dev/null || git merge-base HEAD main 2>/dev/null || git rev-list --max-parents=0 HEAD | tail -n1)" HEAD -- .claude/board/EPIPHANIES.md .claude/board/exec-runs/hhtl-intake-blindness.md

Length of output: 1265

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail

base="$(git merge-base HEAD origin/main 2>/dev/null || git merge-base HEAD main 2>/dev/null || git rev-list --max-parents=0 HEAD | tail -n1)"

printf '%s\n' '== Probe: gate code =='
sed -n '75,310p' crates/helix/examples/probe_hhtl_intake_blindness.rs

printf '%s\n' '== CurveRuler construction =='
sed -n '18,68p' crates/helix/src/curve_ruler.rs

printf '%s\n' '== ResidueEncoder call path =='
sed -n '125,180p' crates/helix/src/residue.rs

printf '%s\n' '== PR diff: probe =='
git diff "$base" HEAD -- crates/helix/examples/probe_hhtl_intake_blindness.rs

printf '%s\n' '== Changed board claims =='
git diff --unified=12 "$base" HEAD -- .claude/board/EPIPHANIES.md .claude/board/exec-runs/hhtl-intake-blindness.md

Length of output: 47325

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail

python - <<'PY'
from collections import defaultdict

def shared_levels(a, b):
    for k in range(4):
        shift = 2 * (4 - k) - 2
        if (a >> shift) != (b >> shift):
            return k
    return 4

def flat_seed(cell):
    return (cell + 4) % 17

groups = {
    "all eligible": lambda a, b: True,
    "(a+b)%7==0": lambda a, b: (a + b) % 7 == 0,
}
for name, keep in groups.items():
    total = agree = 0
    by_k = defaultdict(lambda: [0, 0])
    for a in range(256):
        for b in range(a + 1, 256):
            k = shared_levels(a, b)
            if k == 0 or not keep(a, b):
                continue
            total += 1
            same = flat_seed(a) == flat_seed(b)
            agree += same
            by_k[k][0] += 1
            by_k[k][1] += same
    print(f"{name}: {agree}/{total} = {agree/total:.6%}")
    for k in sorted(by_k):
        n, hits = by_k[k]
        print(f"  shared_levels={k}: {hits}/{n} = {hits/n:.6%}")

print("chance baseline: 1/17 = {:.6%}".format(1/17))
PY

printf '%s\n' '== Current pull-request body =='
gh pr view 876 --repo AdaWorldAPI/lance-graph --json body,url --jq '.url + "\n---\n" + .body' 2>/dev/null || true

Length of output: 6939

---

@AdaWorldAPI I found one remaining H3 reporting defect.

1,152 is not the eligible population. It is the eligible thinned sample after (a + b) % 7 == 0. The full population has 8,064 eligible pairs. The thinning has a small numerical effect here:

• Thinned sample: 51 / 1,152 = 4.4271%
• Full population: 360 / 8,064 = 4.4643%

Please either enumerate all 8,064 eligible pairs or consistently label 1,152 as the deterministic thinned sample. Update the probe output, EPIPHANIES.md, exec-runs/hhtl-intake-blindness.md, and the PR description together. This prevents the corrected H3 denominator from becoming a second population-label error.

H1 now tests the correct constant-shift property and states the API-level argument separately. H2 now has the correct hypothetical-consumer scope.

✅ Action performed

Review finished.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

[AdaWorldAPI]

Confirmed and fixed in 33c9d63 — and you're right that this was the same error one level up, which is the part worth naming.

The codex fix corrected "4,662 = every thinned pair" to "1,152 = eligible", and in doing so I substituted a second population-label error for the first: 1,152 is the eligible thinned sample, never the eligible population. Two denominators, two labels, both wrong, one root cause — a (a+b) % 7 thinning that bought nothing, since 256·255/2 = 32,640 pairs is trivial to enumerate.

So rather than relabel a third time, the thinning is removed. With no sample there is no sample to mislabel, and the class of error is closed rather than patched:

	value
enumeration	exhaustive, all 32,640 cell pairs
eligible population	8,064 pairs sharing ≥1 address level
flat intake	360 / 8,064 = 4.5 %
per-tier intake	8,064 / 8,064 = 100 % by construction

Your independently-computed 360 / 8,064 = 4.4643 % matches exactly. The conclusion is unchanged — 4.5 % sits essentially at the 1/17 ≈ 5.9 % chance level, so the flat intake keeps ancestry about as often as chance would — but it now rests on the whole population instead of a mislabelled sample. Thanks for computing the full enumeration rather than just flagging the label; that's what made the fix obvious.

Updated together so no artifact keeps a stale denominator: probe output, EPIPHANIES.md, exec-runs/hhtl-intake-blindness.md, and the PR description. Both board records carry a dated ⊘ marker recording that the denominator was wrong twice and naming the single root cause — the two-strikes history is more instructive than either correction alone.

Also noted your confirmations that H1 now tests the constant-shift property with the API-level argument stated separately, and that H2 carries the correct hypothetical-consumer scope.

3/3 gates green, clippy -D warnings clean, fmt applied.

---

Generated by Claude Code

[coderabbitai]

Actionable comments posted: 5

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In @.claude/board/EPIPHANIES.md:
- Line 1: Update the H2 entry title in EPIPHANIES.md to remove the withdrawn
claim that the stride-4 suspicion was refuted and shipped constants match a
4-ary consumer. Align the title with the H2 verdict that the suspicion remains
unconfirmed and the 4-ary matching claim was withdrawn, while preserving the
rest of the title’s context.

In @.claude/board/exec-runs/hhtl-intake-blindness.md:
- Around line 21-23: Update the H2 outcome labels in the hypothesis results,
including the corresponding 50-50 entry, from “REFUTED” to “NOT CONFIRMED” or
“suspicion retracted.” Preserve the surrounding explanation and avoid asserting
that the hypothesis was disproven.
- Line 34: Add the language identifier `text` to the code fence on line 34 to
specify that this is a text output block. Update the opening triple-backtick
fence from ``` to ```text to comply with Markdownlint rule MD040, which requires
all code blocks to have a language specified.

In `@crates/helix/examples/probe_hhtl_intake_blindness.rs`:
- Around line 291-305: Clarify the H3 acceptance criterion in the gate around
flat_rate: either derive its threshold from an explicitly declared null model,
or document 0.25 as a hand-tuned smoke-test limit and remove any wording that
treats passing it as evidence of chance-level behavior. Keep the tier_agree
construction check unchanged and ensure the diagnostic message accurately
reflects the selected interpretation.
- Around line 95-158: Add a #[cfg(test)] module in the probe example covering
the helper functions and recorded probe findings: verify the H1 constant shift,
exact H2 prefix_gap_spread outputs, and H3 eligible-pair counts plus agreement
values. Reuse the expected constants and scenarios already defined by the
H1/H2/H3 gates, and keep tests focused on symbols such as start_histogram,
prefix_gap_spread, flat_seed, tiered_seeds, and shared_seed_prefix.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 4575349f-bcbf-4f1e-9bd4-add31c533c09

📥 Commits

Reviewing files that changed from the base of the PR and between c0e42eb and a096114.

📒 Files selected for processing (3)

• .claude/board/EPIPHANIES.md
• .claude/board/exec-runs/hhtl-intake-blindness.md
• crates/helix/examples/probe_hhtl_intake_blindness.rs

[AdaWorldAPI]

All five addressed in d6e6d45. The fourth changed a result, which is the part worth reading.

#4 — the H3 threshold. You're right that flat_rate < 0.25 was hand-tuned and that 24.9 % would have passed it while the prose claimed ≈1/17. I replaced it with a declared null model rather than a better constant: H0 = "the flat seed carries no ancestry information", so its agreement rate among eligible pairs must not exceed its rate among all pairs. Both computed exhaustively.

Running it immediately failed my first formulation — and the failure was correct:

	rate
eligible pairs	360 / 8,064 = 4.4643 %
all pairs (base)	1,800 / 32,640 = 5.5147 %
Δ	−1.05 pp

The rates do not coincide. The flat seed is not ancestry-neutral — it is mildly ancestry-averse: cells sharing an address prefix are numerically close, so they are less likely to be congruent mod 17. My "≈ chance level" phrasing was therefore also wrong, just less wrong than the threshold. The gate now asserts only the true, weaker claim — no ancestry lift — and reports the direction instead of assuming it. Falsifier: an eligible rate materially above base, which would mean the flat seed does carry ancestry.

Worth naming: this is the second time in this probe that replacing a bad criterion revealed the replacement claim was also wrong (the first was the denominator, corrected twice). Both times, making the criterion explicit is what exposed it. A hand-tuned threshold can't fail informatively; a declared null can.

#5 — unit tests. Added a #[cfg(test)] module, 7 tests: the H1 constant shift plus a falsifier that a structure-carrying seed would not be constant-shift (so the H1 test can't pass for any seed); exact H2 gap outputs; the shipped constants including 17 = 4²+1 with the check that stride 4 over 16 covers only 4/16; the H3 population and agreement counts (8,064 / 360 / total); shared-levels as a prefix measure; histogram coverage.

#1 — EPIPHANIES title still carried the withdrawn "REFUTED … matched to 4-ary" conclusion the body had already retracted. Corrected — good catch, that's exactly the failure mode where a corrected record keeps its headline.

#2 — exec-run H2 labels changed REFUTED → NOT CONFIRMED throughout.

#3 — MD040 fence tagged text.

One decline with a reason: clippy stays example-scoped. --all-targets pulls in pre-existing lint debt in helix's own lib tests (manual_is_multiple_of), unrelated to this diff; fixing it here would mix concerns into a probe PR. Scoped run is clean.

3/3 gates green, 7/7 unit tests, clippy clean, fmt applied.

---

Generated by Claude Code