feat(jc): activate Pillar 4 (γ+φ preconditioner) + Pillar 11 (Hambly-Lyons via feature flag)#351
Merged
AdaWorldAPI merged 4 commits intomainfrom May 7, 2026
Merged
Conversation
Replaces the deferred() stub with a concrete probe per the Pillar-4 spec. SOR with ω = GOLDEN_RATIO (1.618) vs vanilla Jacobi (ω = 1.0) on N=50 stiff tridiagonal SPD systems (perturbed 1D Laplacian). Result on default build (no extra features needed): N=50 problems × 16×16, tol = γ/(γ+1)/√N · 1e-6 = 5.176e-8 Jacobi(ω=1): total 12940 iters (mean 258.8) SOR(ω=φ): total 2419 iters (mean 48.4) Step-count ratio = 5.349× (pass if ≥ 2.0×) SOR ≤ Jacobi on 50/50 problems Runtime: ~5 ms The probe uses both: - EULER_GAMMA in the convergence-tolerance scaling factor — same form as lance_graph_planner::cache::lane_eval::NOISE_FLOOR (γ/(γ+1)/√N), tying Pillar 4 to the same Euler-Mascheroni anchor Pillar 5 (Jirak Berry-Esseen) uses for σ-thresholds. - GOLDEN_RATIO as the SOR over-relaxation weight — within ~10 % of theoretical optimal ω* = 2/(1+sin(π/(n+1))) for the 1D- Laplacian-shaped problem class. Both constants from std::f64::consts (stabilized in Rust 1.94). Workspace pinned to 1.94.1 in rust-toolchain.toml — no extra dependencies required. Stiff-regime synthetic problem: diagonal in [2.00, 2.05], off-diagonal in [-1.02, -0.98] — perturbed 1D Laplacian. Spectral radius ρ_J ≈ 0.983 puts the iteration in the regime where over-relaxation decisively wins. (The earlier draft used a diagonally-dominant random SPD problem with ρ_J ≈ 0.4; in that regime optimal ω* ≈ 1.04 and ω = φ over-relaxes — wrong test for the pillar's claim. Stiff regime is the right calibration.) Drops the deferred-pillar count in `prove_it` from 3 to 2 (Cartan-Kuranishi + Hambly-Lyons remain). Hambly-Lyons activates in a follow-up commit under the `hambly-lyons` feature flag. https://claude.ai/code/session_012AUf5NFgeAAQa5aQAKwSgx
…ree-quotient
Replaces the deferred() stub with a concrete probe per the Pillar-11
spec, gated behind the new `hambly-lyons` feature flag. JC's default
build stays zero-dep per the standalone-crate constitution; the
feature flag pulls in the sigker workspace sibling and activates the
probe.
## Probe design (sigker::signature_truncated, depth 2, dim 3)
Two complementary tests over 100 random pairs:
**Forward (tree-equivalence preserves signature):**
Out-and-back path [p₀, p₁, p₀] should have signature ≈ S_identity
by Hambly-Lyons 2010 Theorem 4 (tree-like equivalence collapses to
the constant path's signature).
Result: max ‖S(out-and-back) − S_identity‖ = **0.000e0** across
100/100 pairs (tensor-algebra path is bit-exact at depth-2).
**Converse (non-tree perturbation distinguishes):**
Triangle loop [p₀, p₁, p₂, p₀] has non-zero level-2 signed-area
components.
Result: min ‖S(triangle) − S_identity‖ = **0.0940** across
100/100 pairs (above the 0.05 discrimination threshold).
**Discrimination ratio** (min-converse / max-forward) = **∞**
(perfect separation between tree and non-tree quotient classes at
depth-2 truncation).
## Why the feature-gate (not regular dep, not dev-dep)
JC's constitution: zero EXTERNAL deps in production. sigker is a
workspace sibling (path-dep), not external — but adding it
unconditionally would break the "default cargo build is standalone"
property. dev-dep would put the probe out of reach of run_all_pillars
(dev-deps don't reach lib code). Feature-gated optional dep is the
clean middle path:
cargo run --release --example prove_it
→ 9/11 PASS, 2 deferred (Cartan + Hambly-Lyons)
cargo run --release --features hambly-lyons --example prove_it
→ 10/11 PASS, 1 deferred (Cartan only)
Behind the feature, hambly_lyons::prove() runs the active probe.
Without the feature, it returns DEFERRED with a message telling the
caller how to activate.
## Why this avoids the PR #350 PDE-form correction
PR #350 documents that sigker::kernel::signature_kernel_pde currently
ships a Goursat-PDE form that diverges from the true signature kernel
at moderate inner products. This probe uses signature_truncated (the
tensor-algebra path), which is independent of the PDE-form correction.
Pillar 11 certification holds regardless of #350's outcome.
## Files
- crates/jc/Cargo.toml: sigker as optional dep + hambly-lyons feature
- crates/jc/Cargo.lock: lockfile churn from new optional dep slot
- crates/jc/src/hambly_lyons.rs: replace deferred() with feature-gated
active probe + deferred fallback under cfg(not(feature))
- crates/jc/src/lib.rs: update top docstring (Pillar 4 + Pillar 11
activation history); drop "(DEFERRED)" suffix from Pillar 11
registration label
## Status
Combined with Pillar 4 (commit f4bd6bf — activates unconditionally),
this PR drops the deferred-pillar count from 3 to:
- 2 deferred on default build (Cartan + Hambly-Lyons)
- 1 deferred with --features hambly-lyons (Cartan only)
Cartan-Kuranishi (Pillar 2) remains genuinely deferred — it requires
the learned-attention-mask module from the coupled-revival track,
not a probe-shaped activation.
https://claude.ai/code/session_012AUf5NFgeAAQa5aQAKwSgx
Append-only block documenting the activation of Pillars 4 and 11 (commits f4bd6bf and c191f23 in this branch). Records: - Pillar 4 empirical result (5.349× step-count ratio, 50/50 problems SOR ≤ Jacobi, ~5 ms runtime) - Pillar 11 empirical result (100/100 forward pairs at exactly 0 distance; 100/100 converse pairs above 0.05 threshold; discrimination ratio = ∞) - Architectural rationale for the feature-flag choice (lib code can't use dev-deps; unconditional dep breaks zero-dep constitution; feature-gated optional dep is the clean middle path) - Two stragglers unmissed by the EULER_GAMMA / GOLDEN_RATIO regex pass (non-blocking, in example + workspace-EXCLUDED research crate) - Updated deferred-pillar inventory (3 → 2 default, 3 → 1 with --features hambly-lyons) - Flagged board-hygiene retrofit for PR #348 / #349 as a separate follow-up (Pillar 10 row, sigker types inventory, per-PR archive) https://claude.ai/code/session_012AUf5NFgeAAQa5aQAKwSgx
…hot paths) Per @AdaWorldAPI 2026-05-07: JC's hand-rolled scalar Rust is fine for the proof-in-code role but not for production hot paths. Records the SIMD-critical pillars (substrate / jirak / pearl / dueker_zoubouloglou — all operating on d ∈ {10 000, 16 384}) and the proposed feature-flag shape (`simd-hot` gating optional ndarray dep, mirrors the `hambly-lyons` pattern this PR introduces). Tier-2 pillars (small-dimension: koestenberger, ewa_sandwich, precond, hambly_lyons) explicitly noted as not SIMD-critical — production refactor should target the d=10K-16K loops first. Scope explicitly drawn: SIMD-hot is a separate PR with its own throughput measurements; this PR keeps the Pillar-4 + Pillar-11 activation as a clean atomic deliverable. https://claude.ai/code/session_012AUf5NFgeAAQa5aQAKwSgx
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 4f354a4c08
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".
Comment on lines
+120
to
+122
| for (j, &row_j) in a[i].iter().enumerate() { | ||
| if i != j { | ||
| sigma += row_j * x[j]; |
There was a problem hiding this comment.
When this helper is called with omega == 1.0 for the advertised Jacobi baseline, the loop reads x[j] after earlier rows have already been updated in the same sweep, so the baseline is actually Gauss-Seidel, not Jacobi. This makes the reported Jacobi/SOR step-count ratio and Pillar-4 pass criterion measure the wrong comparison; use a separate previous-iterate vector for the ω=1 baseline or relabel/adjust the proof accordingly.
Useful? React with 👍 / 👎.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Activates two of the three deferred FORMAL-SCAFFOLD pillars in
crates/jc. Drops the deferred-pillar count inprove_itfrom 3 → 2 on default builds and 3 → 1 with--features hambly-lyons. Cartan-Kuranishi (Pillar 2) remains genuinely deferred — it requires the learned-attention-mask module from the coupled-revival track, not a probe-shaped activation.Pillar 4 — γ+φ preconditioner (active on default build)
crates/jc/src/precond.rs— replaces thedeferred()stub with a concrete probe: SOR with ω =GOLDEN_RATIO(1.618) vs vanilla Jacobi (ω = 1.0) on N=50 stiff tridiagonal SPD systems (perturbed 1D Laplacian).Empirical result on default
cargo run --release --example prove_it:Both
EULER_GAMMAandGOLDEN_RATIOfromstd::f64::consts(stabilized in Rust 1.94, workspace pinned to 1.94.1). γ enters as the convergence-tolerance scaling factor — same form aslance_graph_planner::cache::lane_eval::NOISE_FLOOR(γ/(γ+1)/√N), tying Pillar 4 to the Euler-Mascheroni anchor Pillar 5 uses for σ-thresholds.Why stiff regime, not diagonally-dominant: for ρ_J ≈ 0.4, optimal SOR ω* ≈ 1.04 — at ω = φ = 1.618 SOR over-relaxes and slows convergence. The probe deliberately uses the perturbed-1D-Laplacian regime (ρ_J ≈ 0.983), where optimal ω* ≈ 1.690 — within 10% of φ. The earlier draft used the diagonally-dominant regime by mistake; the calibration commit fixes it.
Pillar 11 — Hambly-Lyons signature uniqueness (active under
--features hambly-lyons)crates/jc/src/hambly_lyons.rs— gated behind a newhambly-lyonsfeature flag. Default JC build stays zero-dep; the feature flag pulls insigker(workspace sibling) and runs forward + converse Hambly-Lyons probe viasignature_truncatedat depth 2.Empirical result with
cargo run --release --features hambly-lyons --example prove_it:‖S(out-and-back) − S_identity‖‖S(triangle) − S_identity‖The tensor-algebra path is bit-exact for out-and-back paths at depth-2 truncation — exactly what Hambly-Lyons 2010 Theorem 4 predicts. The signature space cleanly separates tree-quotient classes.
Why feature-gated, not regular dep nor dev-dep:
[dependencies]regular: would unconditionally tie JC's default build tosigker, breaking the "zero-dep production" constitution.[dev-dependencies]: dev-deps don't reach lib code, only tests / examples / benches.hambly_lyons::prove()lives in lib (called byrun_all_pillars).cargo build(default) is zero-dep;cargo build --features hambly-lyonsactivates the probe. Clean middle path.Why this avoids PR #350's PDE-form correction: the probe uses
sigker::signature_truncated(tensor-algebra path), which is independent ofsignature_kernel_pde(the function PR #350 corrects). Pillar 11 certification holds regardless of #350's outcome.Updated deferred-pillar inventory
--features hambly-lyonsOut of scope (flagged as follow-up debt in the ledger)
ndarray::simdfor production hot paths (per @AdaWorldAPI). The hand-rolled scalar math in pillars 1 / 5 / 5b / 8 (operating on d ∈ {10 000, 16 384}) needs to be SIMD-dispatched for production. Same feature-flag shape as this PR'shambly-lyons(a newsimd-hotfeature gating optionalndarraydep). Tier-2 pillars (precond, hambly_lyons, koestenberger, ewa_sandwich) operate on small dimensions and don't gain materially from SIMD. Estimated separate PR: ~150-300 LOC.EULER_GAMMA/GOLDEN_RATIOliteral stragglers the regex pass missed:crates/thinking-engine/examples/codec_rnd_bench.rs:1454-1455andcrates/lance-graph-codec-research/src/zeckbf17.rs:1660. Non-blocking (example + workspace-EXCLUDED research crate). 5-line cleanup PR if you want full hygiene.PR_ARC_INVENTORYarchive entries) — flagged in the ledger; lives in its own follow-up retrofit PR.Test plan
cargo build --manifest-path crates/jc/Cargo.toml --release— clean (default zero-dep build)cargo build --manifest-path crates/jc/Cargo.toml --release --features hambly-lyons— clean (with sigker dep)cargo run --release --example prove_it— 9/11 PASS, 2 deferred (Cartan + Hambly-Lyons fallback)cargo run --release --features hambly-lyons --example prove_it— 10/11 PASS, 1 deferred (Cartan only)precond.rsis the right calibration target (vs the easier diagonally-dominant draft)Files
Commits
f4bd6bfPillar 4 activationc191f23Pillar 11 activation + sigker integration via feature flagb34683dLedger — Pillar 4 + Pillar 11 activation events4f354a4Ledger — JC SIMD follow-up debt (ndarray::simd for hot paths)https://claude.ai/code/session_012AUf5NFgeAAQa5aQAKwSgx
Generated by Claude Code