research: joint-birth noncoalescence, current frontier and independent CPU block

AGENTS.md

@@ -5,16 +5,108 @@ Issue and its latest comments. `docs/STATUS.md` distinguishes current scope
 from the verbatim historical ledger. Old acquisition instructions, opening
 Issue bodies and publication-portfolio snapshots are not a current run queue.
 
-Owner-delegated focus dated 2026-09-13, reconciled with the latest #650:
-- One active acceptance package: consolidate and independently audit #735's
-  arbitrary-period root theorem, with #613/#736's sharp axial full-law boundary,
-  as one probability manuscript. Return the package to #735.
-- Uniform oblique winding corridors are a possible strengthening within this
-  programme, not a new mandatory parallel project or duplicate census.
-- #636 is a visible representation reserve, paused for new extension this round.
-  Reuse #708/#710/#733/#737 rather than re-commissioning their finite results.
-- #275: original-U candidate maps before more sampling; theory-blocked P1,
-  with its original source/normalizer/moving-root contract retained.
+Owner-delegated focus dated 2026-09-29, replacing the Sep13 one-paper allocation:
+- Primary: joint near-critical rank process. #773/#823 supply finite birth/arm
+  inputs; direct jump-two suppression is not a proof against aggregation of two
+  distinct births. This branch now supplies a two-time density bound and an
+  author triangular noncoalescence corollary. A follow-up supplies square
+  NN/matching input transfer and population-IQR/pivotal-clock comparability.
+  Advance continuum torus-observer identification and the joint law, not
+  another request to derive these bounds. No unique copula or universality yet.
+  The new H(s,t) kernel criterion separates rank-only Markov models from
+  entry-history dependence. Exact L3 square/triangular positive/negative
+  controls are available in insertion AND label time. The L512 memory
+  contrast is modest, retrospective and insertion-clock only; do not promote
+  it to continuum certification or redispatch the old 9! enumerations.
+  L3 completion geometry is now computed: triangular direction does not
+  remove memory, whereas a binary completion phase gives an exact four-state
+  count-clock realization. Square unmarked label rank is Markov but marked
+  axial sectors are not. Use the completion-pivotal hazard identity for the
+  next scale-dependent mechanism question; do not redispatch direction-only
+  closure at L3 or confuse count-clock closure with hidden-count label time.
+  A new independent 140k square-L512 completion block is now delivered,
+  fixed at the old count cutoffs. The pooled survival contrast is smaller;
+  exact-D and exact-(D,nu2) weighted finite-lag contrasts are unresolved.
+  The directional kernel's next-step necessity has now been checked:
+  within-D early histories had +1.629 +/- .391 more completion sites in
+  one post-hoc readout. The later independent safe-insertion block gives
+  -.040 +/- .570 and does NOT reproduce it. Do not retain the earlier
+  selected contrast as an accepted empirical exclusion of direction closure.
+  Advance completion-count successor dynamics or a targeted independent
+  prediction, not an automatic next descriptor or reacquisition of this
+  block. Flow-matched Markov completions do not identify actual history.
+  The next-count question is now answered negatively in finite L4: even
+  (rank,D,nu2) fails weak Markovness from the uniform empty start, on both
+  square and triangular lattices. Safe-pair synergy gives the exact
+  two-step transfer and safe-insertion nu drift. Do not redispatch generic
+  finite closure as missing or automatically enumerate the next L.
+  Its large-size importance remains open; use a specified completion-
+  creation prediction, not equality of the tautological immediate hazard.
+  The first large-size safe-insertion block is now delivered: 140k new
+  L512 prefixes, four probes per eligible prefix, primary same-(D,nu)
+  increment contrast +.00668 +/- .01209, unresolved. Only .560% of probes
+  have nonzero increments. Exact projected-component evaluation now resolves
+  the SAME target on all140k prefixes: -.003449 +/- .000772, SE15.65x smaller.
+  This supports finite L512 failure of enlarged (rank,D,nu) closure; it is
+  post-readout estimator refinement, not independent confirmation or a
+  continuum theorem. Direction-only replication remains unresolved.
+  Owner allocation on Sep30: one support worker owns precision/replication;
+  the main investigator advances mechanism, not further precision rounds.
+  The new clock-preserving geometric source leaves all count arrivals and
+  current exit hazard unchanged but has two-step response
+  -s*Var_safe(degree)/(m*(m-1)). Three specified L4 full-lag responses are
+  already calculated; do not redispatch them. Advance finite-window source
+  propagation and the exact safe-path versus survival-weight decomposition
+  in notes/birth-selection-intervention-20260930.md. Replication is not a
+  prerequisite for these derivations or an excuse for another descriptor.
+  The support worker's other independent140k archive readout is now complete:
+  -.00319937 +/- .00062009, same negative sign, no block pooling. Other archive
+  outcomes were already read; this is first e readout, not new prospective
+  production. Do not redispatch that replication or run another precision
+  block. Main's new fixed735-step source response has its own contract at
+  analysis/geometric-source-window-20260930/contract.json; it reuses the first
+  block's permutations to test finite-window transmission, not replication.
+  That735-step job is now completed: susceptibility=-.01104203 +/-.00015767,
+  same55028risk/140000prefixes;14workers,519.145s onTgFr7R, released toReady.
+  Do not rerun it as a missing tool, new precision block or size ladder.
+  The safe-path/selection coupling is now completed too: same1198cells give
+  safe-reference=-.00253706+/-.00063649, natural=-.00220507+/-.00078709,
+  selection=+.00033199+/-.00028738 (unresolved). All140000births match;
+  97666new conditional continuations, not independent preparations. Main next:
+  distinguish realizable entrance and safe-growth sources with nonredundant
+  same-contract readouts. Do not redispatch this coupling or precision.
+  Y and two-step survival at fixed(k,D,c) are affine. Two first-birth-
+  preserving sources also have proportional static M/Etop and moving-root
+  columns at one p; thermal p plus ONE source rank2 does not identify TWO
+  geometric sources. See notes/safe-transport-selection-readout-20260930.md.
+  A separate one-sided-birth source derivation now maps deltaM=deltaEtop
+  through the moving root; symmetricL4 gives a genuine positive normal
+  response and rank2Jacobian. It is a constructed dynamic-source control,
+  not the missing original norm4/U six-coordinate forward columns.
+  Label hazard derivative explicitly contains
+  mean(c)+2mean(e)-Var(c); fixed-count data are not fixed-label observations.
+- Parallel: #809/#833 source-dependent predictive state and noisy information.
+  Reuse #829--#837 and this branch's all-k deterministic four-state encoding
+  theorem, including actual two-row product queries. Distributed deterministic
+  block codes restore rank four. The stochastic parity follow-up retains
+  tensor rank three but weakens single-readout distinction; its product family
+  also has positive initial memory three. Full physical language, transcripts
+  and feedback are separate contracts. Do not infer reliable memory from rank.
+- Bounded writing: #739 already supplies arbitrary-oblique full-law necessity;
+  #771 adds fixed-p consequences. Do not dispatch that proof as missing again.
+- #275: derive normal projected candidate columns under the ORIGINAL source,
+  normalizer and moving root. E_top production existed; missing model maps do
+  not mean that experiment was never run. Aggregate rank is not original U.
+
+These are attention priorities, never task locks or permission classes. The
+owner authorizes useful local/cloud computation; benchmark the specific job
+and use task-local paths. Owner routing update on 2026-09-30: medium and large
+computations go to Huawei first; keep local work for small derivations, scoring
+and editing. This changes compute placement, not scientific permission. Do not
+reset containers or stop unknown jobs. Current
+machine availability must be checked, not inferred from a dated skill snapshot.
+Coordinate through concise repository updates rather than repeated cross-chat
+messages. Keep completed results available even when their lane is not primary.
 
 Choose work by the mathematical or physical distinction it can settle, not by
 how many PRs, scripts, new widths or compatibility scores it generates. A useful