[qmc] frustration-free: symmetric NQS for chiral graviton

tracks/qmc/solutions/frustration-free/challenge-15/README.md

@@ -0,0 +1,162 @@
+# Challenge 15: finite-sphere projected-Pfaffian core
+
+This package implements the auditable core calculation for the finite-size
+lowest-`L=2` sector gap at Laughlin filling `ν=1/3`. It uses one
+sector-conditioned projected-Pfaffian parameter tree for `L=0` and `L=2`.
+
+## Environment
+
+The supported interpreter is CPython 3.12. From this directory:
+
+```bash
+uv sync
+uv run python -c \
+  "import jax; print(jax.default_backend(), jax.config.x64_enabled)"
+uv run pytest -q
+```
+
+`jax.config.x64_enabled` must be `True`. CPU execution is supported. A GPU run
+requires a matching CUDA-enabled `jaxlib`; the runtime provenance in every
+artifact records the backend and x64 state.
+
+## Commands and restart
+
+All subcommands accept `--config PATH` with a JSON object. Command-line values
+override the corresponding JSON fields. The canonical JSON SHA256 is stored in
+each artifact.
+
+```bash
+uv run python -m challenge15.cli oracle \
+  --particles 6 \
+  --output ../../../results/frustration-free/challenge-15/oracle-n6
+
+uv run python -m challenge15.cli train \
+  --particles 6 --ranks 1,2,4 --seeds 0,1,2,3,4 --steps 100 \
+  --output ../../../results/frustration-free/challenge-15/n6
+
+uv run python -m challenge15.cli train \
+  --config n6.json --output ../../../results/frustration-free/challenge-15/n6 \
+  --resume
+
+uv run python -m challenge15.cli evaluate \
+  --checkpoint ../../../results/frustration-free/challenge-15/n6/checkpoint.json \
+  --oracle ../../../results/frustration-free/challenge-15/oracle-n6/result.json \
+  --output ../../../results/frustration-free/challenge-15/n6
+
+uv run python -m challenge15.cli verify \
+  --artifact ../../../results/frustration-free/challenge-15/n6/evaluation.json
+
+uv run python -m challenge15.cli report \
+  --evaluation ../../../results/frustration-free/challenge-15/n6/evaluation.json \
+  --output ../../../results/frustration-free/challenge-15/n6
+
+uv run python -m challenge15.cli manifest \
+  --n6 ../../../results/frustration-free/challenge-15/n6/evaluation.json \
+  --n7 ../../../results/frustration-free/challenge-15/n7/evaluation.json \
+  --n8 ../../../results/frustration-free/challenge-15/n8/evaluation.json \
+  --output ../../../results/frustration-free/challenge-15/final
+```
+
+Training publishes a verified atomic checkpoint after each rank/seed. A resume
+is rejected unless its complete canonical configuration hash matches. Each
+checkpoint includes the shared parameters, Adam state, every PRNG split,
+paired-batch hashes, optimizer diagnostics, source hashes, Git revision,
+runtime versions, and input hashes. Unique sibling partials are file-fsynced,
+verified, atomically replaced, and followed by parent-directory fsync.
+Every load recomputes the training-configuration digest, record coverage,
+serialized-state hashes, completed identities, and rank-parent lineage.
+It also recomputes the execution fingerprint over all package sources,
+`pyproject.toml`, `uv.lock`, Python, JAX, jaxlib, Flax, Optax, NumPy, SciPy,
+SymPy, and h5py versions, x64, backend, platform, and schema policy. Stale code
+or runtime artifacts fail closed.
+
+## Physical and numerical conventions
+
+- `2Q = 3(N-1)`, with north-chart holomorphic monopole spinors.
+- Orbital order follows increasing `2m`; determinant order follows increasing
+  integer bit patterns with ascending creation operators.
+- Distances are physical sphere chords
+  `r_ij = 2 sqrt(Q) l_B sin(gamma_ij/2)`.
+- Energies are in `E_C = e²/(4π ε₀ ε l_B)`.
+- The reported gap is `Δ₂ = E(L=2)-E(L=0)`, not necessarily the absolute
+  lowest neutral excitation.
+- The sparse M=0 low-energy scan separately reports the absolute excitation
+  energy and its integer `L`; it resolves energy-degenerate subspaces with
+  `L²` and fails closed on ambiguous classification.
+- One parameter tree is optimized with `w₀E₀+w₂E₂` on paired sector batches.
+  There are no private sector models.
+
+For `N≤8`, exact coefficient-space evaluation reports energy, overlap, true
+projected-Hamiltonian variance `Var(H_LLL)`, exact angular-momentum residuals,
+quadrature changes, and projected carrier-span singular values. Coordinate VMC
+reports a bare Coulomb potential estimator and its sampling variance. That
+sampling variance is never labeled as `Var(H_LLL)`.
+
+## Fail-closed gates
+
+Ranks are nested as `1,2,4,...`. Two consecutive doublings must each satisfy
+
+```text
+|delta E_L| + 2 sigma_diff <= 1e-4 E_C
+|delta Delta_2| + 2 sigma_diff <= 0.002 Delta_2
+```
+
+Exact rank analysis requires identical seed sets at every adjacent rank and
+uses `sigma_diff=0` by contract. Stochastic analysis is separate: it requires
+at least two identically paired seeds and computes uncertainty from paired
+differences, retaining covariance. Missing pairs fail closed rather than
+returning zero uncertainty. Increased noise makes the left-hand side larger
+and can never relax a gate. Where exact overlaps exist, their change must be at
+most `1e-3`. These tolerances, two required doublings, and the minimum four of
+five seed gate are immutable. Any missing `(rank, seed)`, failed numerical
+gate, or unavailable gate leaves the system explicitly pending.
+
+Smoke optimization uses generated coordinate batches, not an accept/reject
+sampler. Its acceptance-rate fields are therefore JSON `null`; step energies,
+norms, and gradient norms are explicitly labeled as pre-update diagnostics.
+
+The production target-sector solver uses a basis-invariant thin `D×r`
+canonicalization with lexicographic row pivots and only `r×r` gauge matrices;
+it never forms a `D×D` projector. Its immutable symmetry gates are a Gram
+defect at most `1e-12`, an `L²` target residual at most `1e-11`, and a
+reconstructed generator/ladder intertwining residual at most `1e-11`. All
+three diagnostics, pivots, and the linear-storage bound are persisted and
+recomputed when an oracle cache is restored.
+
+Production proceeds in order `N=6`, then `N=7`, then `N=8`; a later size is not
+accepted without a hash- and provenance-bound accepted artifact for the prior
+size. The final manifest remains pending unless all three semantic evaluations
+validate. Full determinant-space dimensions
+are 8,008, 50,388, and 319,770 respectively. Exact construction, sector
+diagonalization, repeated projected-carrier expansion, and doubled quadrature
+dominate memory and wall time. Carrier rank, determinant block size, and
+quadrature order must be reported with peak RSS and elapsed time.
+
+**N=6-8 production acceptance is pending.** Passing every immutable numerical,
+statistical, provenance, and cross-size gate would establish the finite-size
+lowest-L=2 sector gap. It is not a chiral-graviton claim; that requires the
+separate metric-response acceptance plan.
+
+## Audited production controllers
+
+The sole operator entry point is `challenge15 production-orchestrate-size`.
+Low-level Slurm, transfer, deployment, and scientific commands are internal
+create-only state-machine contracts.
+
+- Qdeshell uses `dzagnormal`, account `giggleliu`, QOS
+  `user_jiangweiqi`, and one A800 80 GB GPU per five-way seed array task.
+  The approved result root is
+  `/work/share/giggleliu/jiangweiqi/results/challenge15`.
+- LASG02 uses `ihicnormal`, account `chenkun2025`, QOS
+  `user_student090`, 24 CPUs, and 80000 MiB. The approved result root is
+  `/public/home/student090/results/challenge15`.
+- WUZH02 remains inactive. `discover_wuzh02.py` fails closed unless complete,
+  explicitly labelled scheduler, capacity, Python 3.12, project-root, and
+  result-root evidence is supplied.
+- `xh5-jiangweiqi` is only a documented fallback (account `giggleliu`, QOS
+  `user_jiangweiqi`; available GPU families include 3090, 3080, and V100).
+  No production controller role or exact partition profile is approved by the
+  immutable design, so the orchestrator does not submit there.
+
+Remote `sbatch --test-only`, final deployment receipts, runtime attestations,
+and scientific jobs are deferred until the final clean source freeze.

tracks/qmc/solutions/frustration-free/challenge-15/production/config/base-n4-smoke.json

@@ -0,0 +1 @@
+{"payload":{"burn_in_sweeps":2000,"chains_per_sector":32,"checkpoint_interval_steps":100,"draws_per_update":16,"final_evaluation_burn_in_sweeps":5000,"final_evaluation_chains_per_sector":32,"final_evaluation_draws_per_chain":4096,"final_evaluation_thinning_sweeps":4,"learning_rate":0.001,"optimizer":"adam","pilot_sweeps":500,"reequilibration_sweeps_after_update":4,"refresh_log_amplitudes_after_update":true,"schedule_version":"fixed-v1","steps":10000,"thinning_sweeps":2,"walkers_per_chain":32,"weight_l0":0.5,"weight_l2":0.5},"payload_sha256":"361a7820ae48588f1b7cdbd65f73474aa23a891a7bfddd727178f655ce38043f","schema":"challenge15.production-vmc-config.v1"}

tracks/qmc/solutions/frustration-free/challenge-15/production/config/base-n6.json

@@ -0,0 +1 @@
+{"payload":{"burn_in_sweeps":2000,"chains_per_sector":32,"checkpoint_interval_steps":100,"draws_per_update":16,"final_evaluation_burn_in_sweeps":5000,"final_evaluation_chains_per_sector":32,"final_evaluation_draws_per_chain":4096,"final_evaluation_thinning_sweeps":4,"learning_rate":0.001,"optimizer":"adam","pilot_sweeps":500,"reequilibration_sweeps_after_update":4,"refresh_log_amplitudes_after_update":true,"schedule_version":"fixed-v1","steps":10000,"thinning_sweeps":2,"walkers_per_chain":32,"weight_l0":0.5,"weight_l2":0.5},"payload_sha256":"361a7820ae48588f1b7cdbd65f73474aa23a891a7bfddd727178f655ce38043f","schema":"challenge15.production-vmc-config.v1"}

tracks/qmc/solutions/frustration-free/challenge-15/production/config/base-n7.json

@@ -0,0 +1 @@
+{"payload":{"burn_in_sweeps":2000,"chains_per_sector":32,"checkpoint_interval_steps":100,"draws_per_update":16,"final_evaluation_burn_in_sweeps":5000,"final_evaluation_chains_per_sector":32,"final_evaluation_draws_per_chain":4096,"final_evaluation_thinning_sweeps":4,"learning_rate":0.001,"optimizer":"adam","pilot_sweeps":500,"reequilibration_sweeps_after_update":4,"refresh_log_amplitudes_after_update":true,"schedule_version":"fixed-v1","steps":10000,"thinning_sweeps":2,"walkers_per_chain":32,"weight_l0":0.5,"weight_l2":0.5},"payload_sha256":"361a7820ae48588f1b7cdbd65f73474aa23a891a7bfddd727178f655ce38043f","schema":"challenge15.production-vmc-config.v1"}

tracks/qmc/solutions/frustration-free/challenge-15/production/config/base-n8.json

@@ -0,0 +1 @@
+{"payload":{"burn_in_sweeps":2000,"chains_per_sector":32,"checkpoint_interval_steps":100,"draws_per_update":16,"final_evaluation_burn_in_sweeps":5000,"final_evaluation_chains_per_sector":32,"final_evaluation_draws_per_chain":4096,"final_evaluation_thinning_sweeps":4,"learning_rate":0.001,"optimizer":"adam","pilot_sweeps":500,"reequilibration_sweeps_after_update":4,"refresh_log_amplitudes_after_update":true,"schedule_version":"fixed-v1","steps":10000,"thinning_sweeps":2,"walkers_per_chain":32,"weight_l0":0.5,"weight_l2":0.5},"payload_sha256":"361a7820ae48588f1b7cdbd65f73474aa23a891a7bfddd727178f655ce38043f","schema":"challenge15.production-vmc-config.v1"}

tracks/qmc/solutions/frustration-free/challenge-15/production/deploy/build_source_bundle.sh

@@ -0,0 +1,15 @@
+#!/usr/bin/env bash
+set -euo pipefail
+[[ "$#" == 2 && "$1" == "--output-dir" ]] || { printf '%s\n' "usage: build_source_bundle.sh --output-dir PATH" >&2; exit 2; }
+ROOT="$(cd -- "$(dirname -- "$0")/../.." && pwd -P)"
+OUT="$2"
+[[ "$OUT" = /* && ! -e "$OUT" ]] || { printf '%s\n' "output must be a new absolute path" >&2; exit 2; }
+mkdir -m 700 -- "$OUT"
+(cd "$ROOT" && {
+  printf '%s\n' src production pyproject.toml uv.lock |
+    LC_ALL=C sort > "$OUT/MEMBERS"
+  tar --sort=name --mtime='UTC 1970-01-01' --owner=0 --group=0 --numeric-owner \
+    -cf "$OUT/source.tar" src production pyproject.toml uv.lock
+})
+sha256sum -- "$OUT/source.tar" "$OUT/MEMBERS" | LC_ALL=C sort -k2 > "$OUT/SHA256SUMS"
+sync -f "$OUT/SHA256SUMS"

tracks/qmc/solutions/frustration-free/challenge-15/production/deploy/deploy.sh

@@ -0,0 +1,120 @@
+#!/usr/bin/env bash
+set -euo pipefail
+die() { printf '%s\n' "$*" >&2; exit 2; }
+BUNDLE_ROOT= BUNDLE_SHA256= WHEELHOUSE= REQUIREMENTS= DESTINATION_ROOT= DRY_RUN_RECEIPT= OUTPUT_DIR= PROFILE=
+while (($#)); do
+  case "$1" in
+    --bundle-root|--bundle) BUNDLE_ROOT="$2"; shift 2 ;;
+    --bundle-sha256) BUNDLE_SHA256="$2"; shift 2 ;;
+    --wheelhouse) WHEELHOUSE="$2"; shift 2 ;;
+    --requirements) REQUIREMENTS="$2"; shift 2 ;;
+    --destination-root) DESTINATION_ROOT="$2"; shift 2 ;;
+    --dry-run-receipt) DRY_RUN_RECEIPT="$2"; shift 2 ;;
+    --output-dir) OUTPUT_DIR="$2"; shift 2 ;;
+    --profile) PROFILE="$2"; shift 2 ;;
+    *) die "unexpected argument: $1" ;;
+  esac
+done
+for value in BUNDLE_ROOT BUNDLE_SHA256 WHEELHOUSE REQUIREMENTS DESTINATION_ROOT DRY_RUN_RECEIPT OUTPUT_DIR PROFILE; do
+  [[ "${!value:-}" ]] || die "missing deployment argument: $value"
+done
+[[ -d "$BUNDLE_ROOT" && ! -L "$BUNDLE_ROOT" ]] || die "bundle root must be a regular directory"
+COMPUTED_BUNDLE_SHA="$(
+  cd "$BUNDLE_ROOT"; shopt -s globstar nullglob; files=()
+  for member in **/*; do [[ -f "$member" && ! -L "$member" ]] && files+=("$member"); done
+  printf '%s\0' "${files[@]}" | sort -z | xargs -0 sha256sum -- |
+    sha256sum | awk '{print $1}'
+)"
+[[ "$COMPUTED_BUNDLE_SHA" == "$BUNDLE_SHA256" ]] ||
+  die "bundle SHA256 mismatch"
+(cd -- "$BUNDLE_ROOT" && sha256sum -c -- SHA256SUMS)
+PROFILE_SHA="$(jq -er '.payload_sha256' "$PROFILE")"
+[[ "$PROFILE_SHA" == "$(jq -er '.payload.profile_sha256' "$DRY_RUN_RECEIPT")" ]] ||
+  die "dry-run profile mismatch"
+APPROVED_ROOT="$(jq -er '.payload.approved_project_root' "$PROFILE")"
+[[ "$DESTINATION_ROOT" == "$APPROVED_ROOT" ]] ||
+  die "deployment root is not profile-bound"
+[[ "$DESTINATION_ROOT" == "$(jq -er '.payload.destination' "$DRY_RUN_RECEIPT")" ]] ||
+  die "dry-run destination mismatch"
+[[ "$BUNDLE_SHA256" == "$(jq -er '.payload.bundle_sha256' "$DRY_RUN_RECEIPT")" ]] ||
+  die "dry-run bundle mismatch"
+jq -e '.payload.scheduler_test | length > 0 and
+  all(.[]; .argv[0:2] == ["sbatch","--test-only"] and
+    (.script_sha256 | test("^[0-9a-f]{64}$")))' "$DRY_RUN_RECEIPT" >/dev/null ||
+  die "dry-run scheduler test evidence mismatch"
+DRY_INTERPRETER="$(jq -er '.payload.interpreter' "$DRY_RUN_RECEIPT")"
+[[ -x "$DRY_INTERPRETER" &&
+   "$(sha256sum -- "$DRY_INTERPRETER" | awk '{print $1}')" == "$(jq -er '.payload.interpreter_sha256' "$DRY_RUN_RECEIPT")" ]] ||
+  die "dry-run interpreter tuple is stale"
+FINAL="$DESTINATION_ROOT/$BUNDLE_SHA256"
+RECEIPT="$OUTPUT_DIR/$BUNDLE_SHA256.json"
+validate_existing() {
+  [[ -d "$FINAL" && ! -L "$FINAL" && -x "$FINAL/venv/bin/python" ]] ||
+    die "existing deployment is incomplete"
+  [[ "$(jq -er '.payload.bundle_sha256' "$RECEIPT")" == "$BUNDLE_SHA256" &&
+     "$(jq -er '.payload.profile_sha256' "$RECEIPT")" == "$PROFILE_SHA" &&
+     "$(jq -er '.payload.deployment_root' "$RECEIPT")" == "$FINAL" ]] ||
+    die "existing deployment receipt mismatch"
+  [[ -s "$FINAL/.deployment-sha256s" ]] ||
+    die "existing deployment byte manifest is missing"
+  (cd "$FINAL" && sha256sum -c -- .deployment-sha256s)
+  [[ "$(sha256sum -- "$FINAL/venv/bin/python" | awk '{print $1}')" ==
+     "$(jq -er '.payload.interpreter_sha256' "$RECEIPT")" ]] ||
+    die "existing deployment interpreter changed"
+  [[ "$(sha256sum -- "$FINAL"/.built-wheel/challenge15_nqs-*.whl | awk 'NR==1 {print $1}')" ==
+     "$(jq -er '.payload.installed_wheel_sha256' "$RECEIPT")" ]] ||
+    die "existing deployed wheel changed"
+}
+if [[ -s "$RECEIPT" ]]; then validate_existing; printf '%s\n' "$RECEIPT"; exit 0; fi
+PARTIAL="$DESTINATION_ROOT/.partial.$BUNDLE_SHA256.$(cat /proc/sys/kernel/random/uuid)"
+mkdir -m 700 -p -- "$DESTINATION_ROOT" "$OUTPUT_DIR"
+if [[ -d "$FINAL" && ! -L "$FINAL" ]]; then
+  INTERPRETER="$FINAL/venv/bin/python"
+  [[ -x "$INTERPRETER" ]] || die "promoted deployment is incomplete"
+  INTERPRETER_SHA="$(sha256sum -- "$INTERPRETER" | awk '{print $1}')"
+  INSTALLED_WHEEL_SHA="$(sha256sum -- "$FINAL"/.built-wheel/challenge15_nqs-*.whl | awk 'NR==1 {print $1}')"
+  [[ -s "$FINAL/.deployment-sha256s" ]] ||
+    die "promoted deployment byte manifest is missing"
+  (cd "$FINAL" && sha256sum -c -- .deployment-sha256s)
+else
+mkdir -m 700 -- "$PARTIAL"
+trap 'rm -rf -- "$PARTIAL"' EXIT
+tar -xf "$BUNDLE_ROOT/source.tar" -C "$PARTIAL"
+PYTHON312="$(command -v python3.12 || true)"
+[[ "$PYTHON312" = /* ]] || die "absolute CPython 3.12 interpreter is unavailable"
+"$PYTHON312" -m venv "$PARTIAL/venv"
+INTERPRETER="$PARTIAL/venv/bin/python"
+"$INTERPRETER" -m pip install --no-index --require-hashes --only-binary=:all: \
+  --find-links "$WHEELHOUSE" -r "$REQUIREMENTS"
+mkdir -- "$PARTIAL/.built-wheel"
+"$INTERPRETER" -m pip wheel --no-index --no-deps --no-build-isolation \
+  --wheel-dir "$PARTIAL/.built-wheel" "$PARTIAL"
+PACKAGE_WHEELS=("$PARTIAL"/.built-wheel/challenge15_nqs-*.whl)
+[[ "${#PACKAGE_WHEELS[@]}" == 1 && -f "${PACKAGE_WHEELS[0]}" ]] ||
+  die "deployment did not build exactly one project wheel"
+"$INTERPRETER" -m pip install --no-index --no-deps "${PACKAGE_WHEELS[0]}"
+INSTALLED_WHEEL_SHA="$(sha256sum -- "${PACKAGE_WHEELS[0]}" | awk '{print $1}')"
+INTERPRETER_SHA="$(sha256sum -- "$INTERPRETER" | awk '{print $1}')"
+(cd "$PARTIAL"; shopt -s globstar nullglob; files=()
+ for member in **/*; do
+   [[ -f "$member" && ! -L "$member" && "$member" != .deployment-sha256s ]] &&
+     files+=("$member")
+ done
+ printf '%s\0' "${files[@]}" | sort -z | xargs -0 sha256sum -- > .deployment-sha256s
+ sync -f .deployment-sha256s)
+mv -T -- "$PARTIAL" "$FINAL"
+sync -f "$DESTINATION_ROOT"
+trap - EXIT
+fi
+[[ ! -e "$RECEIPT" ]] || die "deployment receipt exists"
+PAYLOAD="$(jq -Scn --arg dry "$(jq -er '.payload_sha256' "$DRY_RUN_RECEIPT")" \
+  --arg profile "$PROFILE_SHA" --arg bundle "$BUNDLE_SHA256" --arg root "$FINAL" \
+  --arg interpreter "$FINAL/venv/bin/python" --arg ih "$INTERPRETER_SHA" \
+  --arg wheel "$INSTALLED_WHEEL_SHA" \
+  '{dry_run_receipt_sha256:$dry,profile_sha256:$profile,bundle_sha256:$bundle,deployment_root:$root,interpreter:$interpreter,interpreter_sha256:$ih,installed_wheel_sha256:$wheel,deployed_at_utc:(now|todateiso8601)}')"
+PAYLOAD_SHA="$(printf '%s' "$PAYLOAD" | sha256sum | awk '{print $1}')"
+jq -Scn --argjson payload "$PAYLOAD" --arg sha "$PAYLOAD_SHA" \
+  '{schema:"challenge15.deployment-receipt.v1",payload:$payload,payload_sha256:$sha}' > "$RECEIPT"
+sync -f "$RECEIPT"
+sync -f "$OUTPUT_DIR"
+printf '%s\n' "$RECEIPT"