[appleweiping]

Current PhaseSet-V2 upgrade — 2026-09-26

Commit: 093cfc7ac30d1c786b874533e1aba304afa0840a.

Implements signed joint/local complex phase fields, checkpointed temporal
actor–edge incidence, ordered capture readout, common-scale cosine calibration,
complete-relation matching, verified counterfactual loss, and true
mean/difference signal DCT. The new 87-stage bounded experiment matrix
supersedes the affected legacy 33-run plan, while preserving all history.

Actual registered-server checks:

• Focused CPU regression: 117 passed, 1 CUDA-only test skipped (32.47 s).
• Full width512 CUDA forward/backward: 1 passed (7.43 s), RTX A6000;
  peak allocated143108608B, reserved165675008B.
• K2 exact positive-zero topology outputs/parameter gradients, K13/78-edge
  checkpoint/eager equivalence, permutations, masks and relation objectives.
• Source was unchanged in both attempts; private closed receipt digests are
  recorded in the public-safe mechanism report.
• Fresh GPT-5.6-Sol review fixed two fixture defects and found no remaining
  concrete P1/P2; same-family semantic review remains provisional.
• The portable structural journal does not weaken production POSIX durability.

These are independent research APIs, not completed V2 formal-host/baseline
integration or scientific results. Formal stages 0/87, sealed test closed.
Approved Embody main data/body-model assets, exact handoff fold assignments,
and human-verified relational challenge remain outstanding. No toy or
supplemental dataset is promoted to main results; no submission/SOTA claim.

Docs:
amendment,
mechanism report,
87-stage matrix.

Commit-bound CI 36269584021 is all 3 jobs successful.
Both Windows Python3.12/3.14: 1353 passed,26 skipped,39 subtests passed;
test times634.37s/668.67s. Ubuntu public-tree, all-ref, manifest, link, lint,
wheel-content and index-binding audits also passed.

Registered-server full regression from anonymous codeload of this exact commit:
1376 passed,3 skipped,39 subtests passed in727.08s, exit0/source comparison0.
The 18-member downloaded receipt is closed at
308683e36be1f17fa338c20c7ca68f2f3109ab9cf404f396160fcaae2143d7ec.

These close this software mechanism milestone, not native-data training or
scientific acceptance. Previous commit results below remain historical.

---

Preserved previous PhaseSet V1 progress

Current implementation

Latest commit d6d8089275c80ad65ace2af71904ef1960da5487 integrates complete
two-store periodic cache construction/consumption and the explicit cached
residual lifecycle. Both first pass and gradient replay reopen admitted
streams; cached full-capture validation, additive initialization/checkpoint/
report schemas and exact-plan resume preserve the uncached v1 path. Replay
failures restore post-objective RNG before propagating without an optimizer
or scheduler step.

The cache builder enumerates the complete twenty-epoch train census and all
holistic validation windows, admits six registered plans, and reconstructs
the canonical build manifest when consuming an existing tree. Disk limits,
fresh output roots, no fallback and exact source/checkpoint/floor binding are
implemented. This is not an observed licensed-data cache build.

The same commit includes the qualification controller's early POSIX/main-thread
timer check, exception-safe restoration and narrow runtime error classification,
plus an initial portability correction for qualification tests. The current
Windows CI outcome below remains a failure and is not presented as acceptance.

Fifteen explicitly selected files were committed. The exact staged hashes
matched the reviewed inputs; public scanning covered 269 files and 85 local
links, and all-ref scanning covered 14 refs / 501 blobs. Main, the immutable
old tags and published v0.2.1 release remain unchanged.

Actual registered-server evidence

• Combined focused source: 270 passed in 429.82 seconds.
• Same-source full regression: 1360 passed, 2 existing skips, 39 subtests in
  743.26 seconds; source unchanged throughout.
• Actual official CLIP: established two embedding goldens retained, three
  chunks and caller RNG unchanged under the new source binding.

These overlapping software tests use analytic fixtures and are not formal
training, model-quality results or independent experiments. Full-width cached
CUDA equivalence remains a separate uncompleted observation. Its initial
private witness was stopped before execution when review found an invalid
analytic CLIP receipt for the real persisted capture writer; the production
rehydrator is not weakened.

Actions 34286711166
completed with the Ubuntu public-tree/wheel audit passing and both Windows
jobs failing: each reports three failed structural qualification fixtures,
1334 passing tests, 25 skips and 39 subtests. The previous read-only-shard
fixture failure no longer occurs. The remaining failures are being diagnosed;
neither this nor the predecessor CI is retroactively labeled green.

Exact-commit anonymous codeload acceptance passed for all 269 tracked files.
The wheel bound all 71 Python source modules to those bytes. Its isolated
installed-package cohort passed 499 tests in 487.81 seconds, with both source
and installed files unchanged. All 42 receipt members across source, wheel
and installed acceptance were independently downloaded and hash-verified.
These checks do not override the failed Windows jobs.

Remaining work and scientific boundary

Host-level cached configuration and direct plan injection remain separate
implementation work. Private typed cache-build completion evidence passed
109 focused server tests, and private cached checkpoint-ledger binding passed
151; neither is included in this public commit yet. A training-only actor-once
energy-floor collector is under review; supplied floor bytes alone do not
prove scientific fitting.

This is a feature-branch engineering milestone, not v1.0.0 or an empirical
paper release. Licensed native Embody data and the neutral body-model asset
remain unobserved. No nine-base/24-residual formal cohort, qualified winner,
sealed test or empirical aggregate exists. The formal window-training caption
role remains an explicit preregistration clarification. Analytic or supplemental
data do not replace the native main task.

The official-template manuscript remains a no-result draft with technical
content through page four, references on page five, 57 HOLD result cells and
pending author metadata. Same-family semantic review is provisional. Private
data, participants, credentials, endpoints, download links, checkpoints and
operational receipts are excluded. The complete-project goal remains active.