Publish every generator's period - #563
Merged
Merged
Conversation
verify-shipped-analyzers.js built each analyzer project INTO Runtime/Analyzers and diffed the directory against itself before and after. Three paths could not go red (#558): * a renamed or removed Runtime/Analyzers made `before` and `after` both empty maps, so nothing differed and the check reported success forever; * `moved` iterated `after.keys()` only, so a DLL present before and absent after -- the loudest possible mismatch -- contributed no key and was invisible; * a .csproj edit that stopped the copy made the two sides identical by construction, which is the exact scenario the gate exists for, inverted. It now runs the comparison the WallstopProto Generator workflow runs: build each project with AnalyzerPayloadOutputDir redirected at a scratch directory, then compare those bytes against the committed ones over the UNION of both sides. Both shipped assemblies are named rather than discovered, so absence is a failure with a name in it rather than an empty map. Proved red-green against the real repository, not only the fixture: a single appended byte reddens it naming the file, --fix restores it byte-identically and git status is clean afterwards. Self-test registered in run-contract-tests.js with a red half per guard, each asserting its own guard's message so a case cannot pass by tripping a different one. 17 assertions, 0.1 s. Refs #558, #556 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
#556 asked for a meta-check one level up from "no linter has been left unreachable": a linter registered in a suite with no self-test that can make it fail is itself reportable. For a scanner over a corpus that is clean by construction, a green run is not evidence it still fires. test-run-repo-lint.js now asserts every scripts/lint-* and scripts/validate-* has a self-test that INVOKES it and that some registered suite runs. Seven linters are carried on an explicit allowlist naming the issue that tracks them (#562); the check refuses an entry that has since gained coverage, so the list cannot rot in either direction. lint-changelog.ps1 was the eighth and is fixed rather than allowlisted. It read CHANGELOG.md from $PSScriptRoot with no way to hand it anything else, which is why it had no self-test; it now takes -ChangelogPath. Ten red halves, one per rule that exits non-zero, each asserting its own message. A stronger-sounding "the self-test names a red half" predicate was written and DROPPED: all 32 existing self-tests match that vocabulary, so it discriminates nothing -- it would have been one more check that cannot go red, inside the test whose subject is checks that cannot go red. The first draft had #556's own defect. It scanned every file under scripts/tests/ for a mention of the linter, including this file, which names linters in allowlists rather than running them: every allowlisted linter read as covered by the list excusing it. Caught only because the allowlist carries its own red half. Registries are now excluded by name. All four branches proved red on the real tree, then green: a de-allowlisted linter, an allowlist entry that gained coverage, a malformed issue reference, and the clean run. Refs #556, #562 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The owner asked on #516 for "a test or function that ensures that we output the period of each generator (include in our docs) that is ran whenever docs are refreshed". This is that. RandomGeneratorMetadataAttribute gains an optional `period:` argument -- named and defaulted, so no existing annotation or consumer call breaks -- surfaced as RandomGeneratorMetadata.Period and PeriodLabel. All 20 generators declare one, and the Random Generators table carries a Period column that scripts/tests/test-random-periods.js refuses to let drift in either direction: source without docs, docs without source, or two different values. Provenance is part of the value rather than a footnote. A period of 2^128 cannot be observed, so where a specification publishes one it is quoted, and where none exists the value states the MEASURED live state width instead. Four Excellent ratings in this roster came from repositories that now 404; an invented period would be that failure one column over. The measured widths are this session's, taken on 6000.4.6f1 by OR-ing the state delta over 3,000 NextUint draws per generator: 10 bits for WDoomRandom (2^10 = the 1024-draw period its notes already claim, independently corroborated), 251/256 for BlastCircuitRandom, 192/192 for FlurryBurstRandom, 108/160 for IllusionFlow, 28,407 for StormDropRandom. typecheck green in all four configurations. Verified in the real editor through RandomGeneratorMetadataRegistry.Snapshot: 20 generators, 0 undeclared, with the probe gated on a member only the new assembly declares. The gate proved red both ways -- a perturbed docs cell and a removed declaration -- then green. Refs #516, #285 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
PLAN.md gets the session section; .llm/context.md gets three durable rules; unity-api-costs.md gets the measured relational assignment cost and the Debug.LogError trap; unity-devcontainer-testing.md gets the GetTypes() sweep trap. The relational cost went to unity-api-costs.md rather than the MCP skill because that skill reached 480 lines -- exactly the CRITICAL threshold -- and the size gate is a preflight failure, not a warning. Refs #529, #435, #556, #516 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This was referenced Aug 25, 2026
Closed
verify-shipped-analyzers.js exported PROJECTS with a comment saying it was kept for the placement contract test. Nothing imports it: test-analyzer-placement.js enumerates Runtime/Analyzers itself. The export was already dead before this session; the comment I added invented a consumer for it. Found by checking my own justification the way the allowlists in test-run-repo-lint.js require theirs to be checked -- an excuse whose subject does not exist excuses nothing. Refs #558 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
fileInvokes() is in the same file, looks like exactly the right helper, and gives false positives here. Its npm branch expands `npm run <aggregate>`, so validate:local resolves to 31 scripts and every linter inside one reads as covered. Measured: five of the seven allowlisted linters came back "invoked" by test-sync-script-contracts.ps1, which only quotes npm script names in a string literal whose CONTENTS it is asserting. That helper answers "does anything run this" -- the assertion above it -- not "can this go red". A comment-stripped basename token was also measured, to exclude a linter named only in prose. It selects exactly the same seven, so it is complexity with no present effect and was dropped for the same reason the red-half vocabulary predicate was. Refs #556 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The first annotation read "~2^61 expected (Romu capacity estimate)". Romu publishes a CAPACITY -- how much output can safely be drawn -- not a period, and its generators are non-linear with no guaranteed cycle length. In a column headed Period a reader takes 2^61 as one, and I had not verified the figure against the specification either. It now states the two things that are true: no guaranteed period, and the 128/128 state bits measured live this session. Refs #516 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The column now carries entries like "unpublished; 251/256 state bits live (measured)", which is cryptic without the rule behind it. A short section says that every value is a claim and the suffix says whose -- published specification, unverified author claim, or a measurement -- and that a live-state count is a lower bound on state WIDTH rather than a period. Also names where the value lives (the attribute on the type) and that a contract test fails the build if the table and the annotation disagree. Refs #516 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Eli Pinkerton (wallstop)
changed the title
Two small things on the shipped surface. The attribute declared Period between Quality and Notes, splitting the run of constructor-ordered properties; the struct next door already declares it after ReferenceUrl. It now matches both the struct and the constructor. Its <remarks> named scripts/tests/test-random-periods.js. That path is real but it is contributor tooling, and this doc shows up in a consumer's IntelliSense. The remark now explains the rule -- every value says whose claim it is -- and leaves the enforcement to the contributor docs that already carry it. typecheck green in all four configurations. Refs #516 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Collaborator
Author
|
Note for review, correcting one of my own commit messages.
Nothing durable was lost — the session's findings are in the tracked files ( Same class as the two claims this PR already corrects in its own code: a dead |
The existing note said superseding is "pure waste" and reds the previous commit. Both true, and both easy to reason around the wrong way: this session pushed four times, checking each time that the Unity legs were still QUEUED and concluding nothing was being wasted. That measured the wrong thing. require-current-pr-head fails every leg of the superseded run as stale whether or not those legs ever started, so three red Unity Tests runs are now on this branch and queue depth had nothing to do with it. The note also now says what does NOT break, because I nearly wrote the scarier version: `Unity CI Success` re-resolves the pull request head last and reported success on all three superseded runs. That is deliberate -- the job carries the same measurement in its own comment -- so the damage is a misleading run list, not a blocked gate. Saying otherwise would push the next session into being over-cautious about a red that does not gate anything. Held rather than pushed: pushing this would have made it four. Refs #435 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The standalone leg reddened on ColorQuantizationTests' fuzz test:
Cutoff 0.8862745 disagreed at channel 226. Expected: False But was: True
Real bug, not a flake. ToThresholdByte returned floor(cutoff * 255f), which is
the ALGEBRAIC inverse of the decode and not the float one -- and this class's
own remarks quote the rule it breaks: never mix the encode and decode steps of
two quantizers. ToNormalized multiplies by the rounded reciprocal 1f/255f;
the threshold multiplied by 255f. Within a ULP of a boundary the multiply by
255 recovers exactly the value the reciprocal lost, so the two disagree.
Measured on 6000.4.6f1 at +/-2 ULP of every channel boundary before the fix:
126 disagreements across 3,820 probes. The fuzz test could only ever find one
of them, and only by luck -- it draws cutoffs from an unseeded PRNG.Instance.
The answer is now defined against ToNormalized itself, which is the comparison
the caller actually makes: seed with the fast floor, then correct by at most
one step in each direction. After: 457,216 probes at +/-3 ULP of every
boundary against every channel, 0 disagreements, saturation unchanged
(NaN 0, -inf 0, +inf 255).
Two deterministic tests replace luck with a walk: every boundary at -1/0/+1
ULP against its neighbouring channels, and the exact pair CI reported.
ColorQuantizationTests 12 pass / 0 fail through the editor, gated on a member
only the fixed assembly declares.
ColorQuantization is not in 3.5.1 (`git ls-tree` on the tag), so this folds
into its Added entry rather than becoming a Fixed one.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Why: No generator stated its period anywhere a caller could read it, and two of the gates that were supposed to catch drift could not fail.
What:
RandomGeneratorMetadata.Period, annotate all 20 generators, and gate the docs table against it so the two cannot drift.lint-changelog.ps1a-ChangelogPathand ten red halves; it could only ever read one file, which is why it had no self-test.Fixes #558
Note
Medium Risk
Touches runtime color threshold semantics and documents RNG period claims consumers may rely on; most other changes are additive metadata and CI/tooling, with moderate blast radius if gates regress.
Overview
Adds
RandomGeneratorMetadata.Period(plusPeriodLabel) and annotates every shippedIRandomimplementation with either a published period or a measured state-width claim. The random-generators docs table gains a Period column, andtest-random-periods.jsfails the build if any generator omits a period or if docs and attributes disagree.ColorQuantization.ToThresholdByteno longer uses a barefloor(cutoff * 255f); it adjusts the candidate so inclusive comparisons againstToNormalizedstay correct at float boundaries, with boundary-focused tests including the CI regression case.Repo hygiene and gates:
verify-shipped-analyzers.jsnow builds into a scratch dir and byte-compares namedANALYZERSDLLs toRuntime/Analyzers(missing dir, missing/stale/unexpected DLLs are explicit failures), withtest-verify-shipped-analyzers.js.lint-changelog.ps1accepts-ChangelogPathand gains a full self-test;test-run-repo-lint.jsadds a meta-check that every linter has a reachable self-test (registry files excluded; seven linters still tracked on #562). Contract tests andpackage.jsonwire the new scripts..llmskills/context record related traps (V8 regex, linter coverage scans, Unity test namespace filtering,Debug.LogErrorcost on relational misses).Reviewed by Cursor Bugbot for commit 3ce5c95. Bugbot is set up for automated code reviews on this repo. Configure here.