[DreamingOfClouds]

Description

What this is

research/p104-designer-toolkit-v2 packages the designer tooling behind the recent P104 frontier line as one command, tk, with a regression suite. It claims no kappa and no new word. The frozen predecessor research/p104-designer-toolkit is untouched; this package supersedes it (same exact-pricing core, plus everything below). Placement follows the repository convention (research/<slug> with README.md, TASK.json, SOURCES.json, VERIFY.json); it is registered through VERIFY.json.

What the toolkit does

• Exact pricing and legality. tk price gives the exact invoice and Decimal-50 crossing of a word with side constraints and budgets; tk check lists every legality violation localised to rule, event and register role (fast oracle, exact on the checked rules); tk accept runs the PR's own price.run with cached native pairs; tk adopt derives a verifier profile from a PR head (profiles/*.json, with the PR's own local.py/pricing.py kept unmodified).
• Rebuild pipeline. tk rebuild WORD [--sides skeleton,routing,hubs,markets,signal] [--post fold,hoist,retain] re-derives the skeleton, Y routing, hubs, markets and signal events of Conditional κ .003125804 (+5.22e-7 over #5907): remove helper 470's frame-593 rung, the twin of #5907's removal #6030 from pinned rule builders (alignbuild/), each step accepted only at exact kappa (never-lose). --post fold,hoist,retain then applies the constructor passes behind [Superseded by #6144] Conditional κ .003127137 (+1.333e-6 over #6030): owned registers 770 and 771 descend through the frames of helpers 705 and 706, which no longer relay them #6143/Conditional κ .003128386 (+2.582e-6 over #6030): owned registers 770, 771 and 772 descend through the frames of helpers 705, 706 and 731, which no longer relay them; helper 732 reads 771 at frame 117 #6144 (R-FOLD shadow-relay fold, H-HOIST gauge stop drop) and Conditional κ = 0.003128645 (+2.59e-7 over #6144): retain an operand to remove a root stop #6167 (R-RETAIN retained-operand stop drop). From a clean clone, tk rebuild 6030 --post fold,hoist,retain runs in about 11 minutes (rebuild 2 min, fold 9 min for 3 accepted sites, hoist and retain under 1 s) and ends at kappa .003128652015904059, the exact kappa of Conditional κ = 0.003128645 (+2.59e-7 over #6144): retain an operand to remove a root stop #6167. The intermediate kappas are .003127144482523 after the second fold site (= [Superseded by #6144] Conditional κ .003127137 (+1.333e-6 over #6030): owned registers 770 and 771 descend through the frames of helpers 705 and 706, which no longer relay them #6143), .003128393203271 after fold,hoist (= Conditional κ .003128386 (+2.582e-6 over #6030): owned registers 770, 771 and 772 descend through the frames of helpers 705, 706 and 731, which no longer relay them; helper 732 reads 771 at frame 117 #6144) and .003128652015904 after retain (= Conditional κ = 0.003128645 (+2.59e-7 over #6144): retain an operand to remove a root stop #6167). retain applied to the shipped Conditional κ .003128386 (+2.582e-6 over #6030): owned registers 770, 771 and 772 descend through the frames of helpers 705, 706 and 731, which no longer relay them; helper 732 reads 771 at frame 117 #6144 word reproduces Conditional κ = 0.003128645 (+2.59e-7 over #6144): retain an operand to remove a root stop #6167's program bit for bit (tests/retain_test.py).
• Generators. tk gen2 (markets, Y-route, late stages, never-lose stage pipeline genpipe2.py), tk gen3 (gen2 plus primgen, signal and retain stages, genpipe3.py), tk synth / tk gensearch / tk genpipe (spec-space generators and searches).
• Sub-circuit re-solve tools. subcirc/ (implementation-free I/O specs of the market columns of Conditional κ .003125804 (+5.22e-7 over #5907): remove helper 470's frame-593 rung, the twin of #5907's removal #6030, variant library), surgery/ (extract and splice of octet columns with a bit-exact round trip and SPAN5/JR validation), tk anatomy, tk bulk (whole-corpus table), tk shipcheck.
• Move classes and screens. Exact CP-SAT retimes (retime, evretime, reorder, reframe), hand-off, helper-split, primitive-basis and macro screens, mined windows, and a C++ SA kernel (k2/).
• Floor LPs and theory. infra/floors_true.py, model3.py, linwt.py, rigidity.py, with P104-THEORY.md (structure of the frontier word and the floor arguments) and P104-LEMMAS.md (what is proved, refuted and conditional, gaps listed).

How to run it

python3 -m pip install numpy scipy sympy       # ortools only for the CP-SAT commands
cd research/p104-designer-toolkit-v2
python3 -B tk.py --help
python3 -B tk.py price 6030                    # exact kappa .003125811557911897
python3 -B tk.py rebuild 6030 --post fold,hoist,retain --workdir /path/to/work
python3 -B tests/regress.py --fast             # about 80-160 s
python3 -B tests/regress.py --quick            # the full stdlib-fast suite (about 15 minutes)

README.md has the quickstart and per-command notes; MAP.md the command reference and module map. Heavy commands run synchronously by default; TK_HQ=/path/to/wrapper routes them through an optional job queue.

Evidence

• scripts/research_package.py --package research/p104-designer-toolkit-v2 replays VERIFY.json (tests/regress.py --fast plus tests/retain_test.py) on Python 3.12 with numpy 2.2.6, scipy 1.14.1 and sympy 1.13.3 (the generic CI pins): PASS_REGISTERED_RESEARCH_PACKAGE; --check-only and check_research_commit.py --check-modes pass; tests/test_research_*.py (702 tests) pass.
• --fast: 33 checks, 0 failed (exact kappa of PR5545/5551/5553/5561/5561s/5907/5827, native replay equality, localised legality mutants, screens executed on Improve conditional κ to 0.00310676333 with a four-gate contact rewrite #5660, shipcheck, surgery round trip on Conditional κ .003125804 (+5.22e-7 over #5907): remove helper 470's frame-593 rung, the twin of #5907's removal #6030, rebuild skeleton+hubs mini chain at exact kappa).
• --quick: 104 checks, 0 failed on a clean git archive export of the final commit (Python 3.12, numpy 2.2.6, scipy 1.14.1, sympy 1.13.3, mpmath 1.3.0; no networkx, no ortools, no private inputs; the SKIP lines above apply). It additionally runs admit, incr/nincr, prims/macros/ctxmine, generators (FRONTGEN, MAINHUB, HUBGEN), accept of the adopted profiles, surgery and bulk tests.
• tk rebuild 6030 --sides skeleton,hubs: 37/37 steps at exact kappa .003125811557911897 (about 1 minute).

Honest scope and limitations

• Tooling only. Models are not oracles: retime/evretime/reorder objectives are linearised kappa and the reported kappa is always the exact re-price. Free helper start/final frames and the owned-inventory pins are verifier conventions, not maintainer-approved contract.
• The rebuilt words are legal at the target kappa but are not program-identical to the swarm's hand-built words (for example 23346 events for the Conditional κ = 0.003128645 (+2.59e-7 over #6144): retain an operand to remove a root stop #6167-kappa word against 22821 in Conditional κ = 0.003128645 (+2.59e-7 over #6144): retain an operand to remove a root stop #6167). What is reproduced exactly is kappa, the pass sequence and the single-site edits (bit-identical for retain).
• rebuild, surgery and anatomy carry Conditional κ .003125804 (+5.22e-7 over #5907): remove helper 470's frame-593 rung, the twin of #5907's removal #6030's register/frame numbering (alignbuild/data/); other words must share it or bring their own manifest. Steps that do not solve are logged GAP and skipped (one such step in the Conditional κ .003125804 (+5.22e-7 over #5907): remove helper 470's frame-593 rung, the twin of #5907's removal #6030 chain).
• Not runnable from a clean clone because inputs are not shipped: tk gen (S4GEN trees), planbuild/plan.py and build.py (plan-search model trees; shipped as source only), hops --recomb, the BIT-side drivers that need adopted BIT packages, and the k2 engine tests that need unshipped nets (k2unify_test, k2levers_test print SKIP). tk adopt needs network access. leverbank_test/genpipe2_test need networkx, which generic CI lacks (they print SKIP there; the gen2 generators themselves need it). The CP-SAT legs need ortools (--solver).
• Behavioural fix made while packaging: the rebuild builders' intermediate words are now judged under the input word's verifier profile (fold/hoist) and the newest profile of the family (retain, which adds a lifetime). Before this, once a Conditional κ = 0.003128645 (+2.59e-7 over #6144): retain an operand to remove a root stop #6167 profile existed, the unmodified chain judged every step under the 32-lifetime profile and logged 37/37 GAP. See alignbuild/api.py, rebuild.py, alignbuild/hoist.py, alignbuild/fold.py.
• Other packaging changes: absolute paths replaced by package-relative ones; scratch state goes to TK_WORK (default: the package directory, git-ignored); a few tests that depended on private inputs were gated or rewritten (framegap uses a synthetic bad word, the speed test threshold counts the failing pairs a fresh cache has); helper modules that lived outside the toolkit were copied in (alignbuild/fold.py, iob2.py, foldcensus.py, subcirc/pdata.py, surgery/cen.py, surgery/s5lib.py, ledger.py). Unmodified swarm verifier sources are byte-pinned in SOURCES.json.

Credits

Built on the swarm's published work: the #6030 lineage (#6030, #6143, #6144) whose constructor steps fold/hoist reproduce; #5907 and #5827 (profile words and tests); #6167 (aliromman), whose retained-operand stop drop is retain; #6163 (petalas) and #6166 (Matarener); and the earlier P104 line (#5545, #5551, #5553, #5561, #5599, #5612, #5660) whose verifiers are vendored unmodified. Their headers and licenses are kept.

Open next questions

1. Can the unshipped model inputs of planbuild/ be packaged so plan.py/build.py run from a fresh clone? First step: list the modules they import and vendor the minimal closure.
2. Can fold (about 3 minutes per accepted site) be made fast enough for generic CI to replay Conditional κ .003125804 (+5.22e-7 over #5907): remove helper 470's frame-593 rung, the twin of #5907's removal #6030 -> Conditional κ .003128386 (+2.582e-6 over #6030): owned registers 770, 771 and 772 descend through the frames of helpers 705, 706 and 731, which no longer relay them; helper 732 reads 771 at frame 117 #6144 end to end? First step: profile alignbuild/iob2.py on one site and cache the rule matrices.
3. Are free helper start/final frames and the owned-inventory pins part of the caller contract? Blocked on maintainer review; first step: tk retime 5551 --free starts against tk diff.

Prepared by DreamingOfClouds with Claude Opus 5.5 (Anthropic), via Claude Code.

🤖 Generated with Claude Code