feat(planner)!: decouple source and opening commitment bases

.github/workflows/ci.yml

@@ -153,7 +153,7 @@ jobs:
           # --lib --bins --tests matches the default test target set but skips
           # compiling examples (clippy's --all-targets already compile-checks them).
           cargo nextest run --lib --bins --tests --profile ci --cargo-profile ci-test --no-default-features --features parallel,disk-persistence,schedules-default \
-            --partition "slice:${SHARD_INDEX}/${SHARD_TOTAL}"
+            --partition "count:${SHARD_INDEX}/${SHARD_TOTAL}"
           status=$?
           end_epoch=$(date +%s)
           if [ -f target/nextest/ci/junit.xml ]; then
@@ -187,8 +187,13 @@ jobs:
           shared-key: ci-test-all-schedules-drift
           add-job-id-key: false
           cache-on-failure: true
+      - name: Committed catalog coverage guard
+        run: |
+          cargo test --release -p akita-config --features all-schedules generated_catalogs_cover_emitted_keys
       - name: Generate schedule tables
         run: scripts/generate-schedule-tables.sh
+      - name: Committed schedule tables match the generator
+        run: git diff --exit-code -- crates/akita-schedules/src/generated
       - name: Drift guard (all-schedules)
         run: |
           cargo test --release -p akita-config --features all-schedules generated_schedule_tables_match_key_planner

.github/workflows/profile-bench.yml

@@ -28,9 +28,9 @@ env:
   #   and distributed profiles use the corresponding adaptive catalogs.
   AKITA_BENCH_MODE: onehot_fp128
   AKITA_BENCH_RUNS: "3"
-  # One warm-up run per case is discarded before the measured runs to
-  # prime caches and lazily-initialized statics; cache misses on the
-  # first run otherwise distort the median across only 3 samples.
+  # One warmup process per case is discarded before the measured processes.
+  # This lets the runner reach a steady state and warms operating system file
+  # caches. Each measured process constructs fresh setup and local state.
   AKITA_BENCH_WARMUPS: "1"
   # The profile example emits the per-fold wire grind_nonce in its proof-size
   # breakdown; scripts/profile_bench_report.py surfaces it per level.
@@ -53,33 +53,40 @@ jobs:
         # deterministically, fixing case order in the comment. `cases` is a
         # newline-separated list of MODE:NUM_VARS:NUM_POLYS[:SETUP_MODE] specs.
         #
-        # The small-field (fp32/fp64) D128 one-hot cells run at nv=28, not nv=32.
-        # Their ext-degree-4 challenge schedule keeps a much larger un-folded witness
-        # than the fp128 cells (the fp32 nv=32 setup vector is ~2.6 GiB vs ~270 MiB
-        # for fp128), so at nv>=30 the prover's eq-evaluation table exceeds the
-        # 1 GiB `MAX_MATERIALIZED_EQ_TABLE_BYTES` ceiling and the run aborts. nv=28
-        # is the largest validated size that both fits that budget and still folds
-        # through multiple recursive levels. The fp128 cells fold aggressively enough
-        # to stay at nv=32.
+        # Keep each base field in its own feature graph. This avoids compiling
+        # all three fields and their schedule catalogs into every base-profile
+        # runner. The small-field one-hot cells use the largest rows shipped by
+        # this stack; the adaptive-dimension follow-up raises them to nv=30.
         group:
-          - name: 1-fp128-dense
-            cases: |
-              dense_fp128:24:1
-          - name: 3-flat-onehot-suite
+          - name: 1-fp32-base
+            profile_feature: profile-ci-fp32
             cases: |
+              dense_fp32_d128:26:1
               onehot_fp32_d128:28:1
+          - name: 2-fp64-base
+            profile_feature: profile-ci-fp64
+            cases: |
+              dense_fp64_d128:26:1
               onehot_fp64_d128:28:1
+          - name: 3-fp128-base
+            profile_feature: profile-ci-fp128-base
+            cases: |
+              dense_fp128:26:1
               onehot_fp128:32:1
           - name: 4-multi-group-direct
+            profile_feature: profile-ci-multi-group-direct
             cases: |
               onehot_fp128_multi_group:32:4:direct
           - name: 5-multi-group-recursive
+            profile_feature: profile-ci-multi-group-recursive
             cases: |
               onehot_fp128_multi_group_recursive:32:4:recursive
           - name: 6-multi-group-recursive-w8r2
+            profile_feature: profile-ci-multi-group-recursive-w8r2
             cases: |
               onehot_fp128_multi_group_recursive_multi_chunk_w8r2:32:4:recursive
           - name: 7-distributed
+            profile_feature: profile-ci-distributed
             cases: |
               onehot_fp128_multi_chunk_w2r2:32:1
               onehot_fp128_multi_chunk_w4r2:32:1
@@ -92,8 +99,12 @@ jobs:
       - name: Initialize benchmark paths
         run: |
           group_dir="$RUNNER_TEMP/bench/${{ matrix.group.name }}"
+          profile_threads="$(nproc)"
+          if [ "$profile_threads" -lt 2 ]; then
+            echo "Profile CI requires at least two CPUs for the multi threaded verifier timing" >&2
+            exit 1
+          fi
           echo "AKITA_BENCH_GROUP_DIR=$group_dir" >> "$GITHUB_ENV"
-          echo "LOCALAPPDATA=$group_dir/setup-cache" >> "$GITHUB_ENV"
           # pr/ holds the PR-side summary, base/ the merge-base side; both go
           # into one per-group handoff artifact for the report job.
           echo "AKITA_BENCH_ARTIFACT_DIR=$group_dir/pr" >> "$GITHUB_ENV"
@@ -102,6 +113,8 @@ jobs:
           echo "AKITA_BENCH_SOURCE_BRANCH=${{ github.head_ref || github.ref_name }}" >> "$GITHUB_ENV"
           echo "AKITA_BENCH_BASE_REF=" >> "$GITHUB_ENV"
           echo "AKITA_BENCH_MERGE_BASE_SHA=" >> "$GITHUB_ENV"
+          echo "AKITA_PROFILE_PROVE_THREADS=$profile_threads" >> "$GITHUB_ENV"
+          echo "AKITA_PROFILE_VERIFY_THREADS=$profile_threads" >> "$GITHUB_ENV"
 
       - name: Determine PR benchmark merge base
         if: github.event_name == 'pull_request'
@@ -167,35 +180,29 @@ jobs:
           lscpu || head -n 40 /proc/cpuinfo
           echo "::endgroup::"
 
-      # PR deltas must compare binaries measured on this runner: ubuntu-latest
-      # is a heterogeneous fleet, and baselines measured on different silicon
-      # produced phantom double-digit deltas.
-      - name: Verify profile-ci feature coverage
+      # PR deltas compare binaries measured on the same custom runner so
+      # hardware differences cannot create false changes.
+      - name: Verify profile benchmark feature coverage
         run: ./scripts/check_profile_ci_features.sh
 
-      - name: Generate schedule tables
-        run: scripts/generate-schedule-tables.sh
-
       - name: Build merge-base profile binary
         if: github.event_name == 'pull_request' && env.AKITA_BENCH_BASELINE_MODES_OK == '1'
         env:
-          AKITA_BENCH_GROUP_CASES: ${{ matrix.group.cases }}
+          AKITA_BENCH_PROFILE_FEATURE: ${{ matrix.group.profile_feature }}
         run: |
           git worktree add "$RUNNER_TEMP/bench-base" "$AKITA_BENCH_MERGE_BASE_SHA"
           mkdir -p "$RUNNER_TEMP/bench-bins"
-          features="parallel,profile-ci"
-          if grep -q ':recursive' <<< "$AKITA_BENCH_GROUP_CASES"; then
-            features="$features,disk-persistence"
-          fi
+          features="parallel,$AKITA_BENCH_PROFILE_FEATURE"
           (
             cd "$RUNNER_TEMP/bench-base" &&
-            if [ -x ./scripts/generate-schedule-tables.sh ]; then
-              ./scripts/generate-schedule-tables.sh
-            fi
-            if python3 scripts/cargo_feature_exists.py akita-pcs profile-ci; then
+            if python3 scripts/cargo_feature_exists.py akita-pcs "$AKITA_BENCH_PROFILE_FEATURE"; then
               CARGO_TARGET_DIR="$RUNNER_TEMP/bench-base-target" \
                 cargo build --release --quiet --example profile \
                   --no-default-features --features "$features"
+            elif python3 scripts/cargo_feature_exists.py akita-pcs profile-ci; then
+              CARGO_TARGET_DIR="$RUNNER_TEMP/bench-base-target" \
+                cargo build --release --quiet --example profile \
+                  --no-default-features --features "parallel,profile-ci"
             else
               CARGO_TARGET_DIR="$RUNNER_TEMP/bench-base-target" \
                 cargo build --release --quiet --example profile
@@ -224,17 +231,17 @@ jobs:
 
       - name: Build profile binary
         env:
-          AKITA_BENCH_GROUP_CASES: ${{ matrix.group.cases }}
+          AKITA_BENCH_PROFILE_FEATURE: ${{ matrix.group.profile_feature }}
         run: |
-          features="parallel,profile-ci"
-          if grep -q ':recursive' <<< "$AKITA_BENCH_GROUP_CASES"; then
-            features="$features,disk-persistence"
-          fi
+          features="parallel,$AKITA_BENCH_PROFILE_FEATURE"
           cargo build --release --quiet --example profile \
             --no-default-features --features "$features"
 
-      - name: Verify profile-ci linkage isolation
-        run: ./scripts/check_profile_ci_linkage.sh target/release/examples/profile
+      - name: Verify profile linkage isolation
+        run: >-
+          ./scripts/check_profile_ci_linkage.sh
+          target/release/examples/profile
+          ${{ matrix.group.profile_feature }}
 
       # On PRs the head and merge-base binaries run interleaved (PR, base,
       # PR, base, ...) so machine-state drift on shared runners lands on both

book/src/foundations/spec-index.md

@@ -23,7 +23,13 @@ quantum LGSA policy at a 128-bit target.
 Recent archived records include
 [`group-local-opening-points`](../../../specs/archive/2026-Q3/group-local-opening-points.md),
 whose durable claim ownership and protocol dataflow now live in the architecture,
-verification, commitment API, and extension-opening chapters.
+verification, commitment API, and extension-opening chapters. The
+[`PR 375 prover optimization record`](../../../specs/archive/2026-Q3/pr375-prover-streaming-and-onehot-unification.md)
+now lives in the same archive. Its durable source ownership, CPU resource, and
+NTT lifecycle rules live in the optimization and commitment API chapters. The
+[`profile-bench-coverage-matrix`](../../../specs/archive/2026-Q3/profile-bench-coverage-matrix.md),
+whose current benchmark contract now lives in the profiling chapter, is also
+archived.
 
 ## Sources to fold in
 

book/src/how/configuration.md

@@ -53,4 +53,4 @@ owns shipped table data. The verifier-reachable proof-size formula.
 - `crates/akita-types/src/proof_size.rs` and `crates/akita-types/src/layout/proof_size.rs` (`level_proof_bytes`, planned witness sizing).
 - Paper §3.11 `sec:akita-planner` (objective/constraints, the dynamic program, generated schedules).
 - `crates/akita-config/src/generated_families.rs`, `crates/akita-schedules/src/generated/`, `crates/akita-planner/src/resolve.rs` (`resolve_schedule`).
-- `book/src/usage/profiling.md`, `specs/profile-bench-coverage-matrix.md`, `.github/workflows/profile-bench.yml`.
+- `book/src/usage/profiling.md` and `.github/workflows/profile-bench.yml`.