ci: build the playground with the real emcc on every PR; [99i3] names itself an approximation (#1255) - #1267
Merged
Merged
Conversation
… itself an approximation (#1255) pages.yml ran only on push to main, so the wasm32 build of web/build.sh was first attempted after a merge and stayed red on main for five commits while [99i3] (clang -m32) was green: i386 aligns double to 4 inside a struct and wasm32 to 8, so a layout _Static_assert can pass -m32 and fail emcc. - pages.yml: the build job also runs on pull_request (same path filter, same `bash web/build.sh`, no second copy of the emcc flags). Configure Pages, the artifact upload and deploy are gated to non-PR events; a PR gets its own concurrency group so it cannot cancel a main deploy. - tools/ilp32_syntax_check.sh: with emcc on PATH, every recorded TU is compiled by the real emcc (-fsyntax-only, the call's recorded flags plus its -s settings) and the verdict is AUTHORITATIVE; otherwise the -m32 arm runs and the verdict/OK lines say APPROXIMATION and name pages.yml as the authority. Population, examined == len > 0, floor and all plants unchanged. New self-test control 1w (the -m32 arm passes a 12-vs-16 layout assert) and plant 1wt (the wasm32 frontend refuses it) pin the limit from both sides; with emcc on PATH the real arm is also driven and must be red. - [99i3]: label and comment no longer claim the build "cannot break unnoticed"; requires the verdict: line and reads OK/macro_parity for that mode; case count 55 -> 57. - werror_switch_check: ilp32 script floor 12 -> 14 (the real-emcc compile line and plant 1wt, measured with --print-counts). - docs/CI.md (new section + check row), docs/PERFORMANCE.md, CHANGELOG; docs_claims_populations PATHS|docs/CI.md 53 -> 63. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…ayground check reports on every PR Gap 1 (oracle): round 1's real-emcc arm compiled each TU with -Isrc, -DEIGENSCRIPT_EXT_*=0, -DEIGENSCRIPT_VERSION and -Werror=* that web/build.sh never passes, so a recipe the driver rejects printed `verdict: AUTHORITATIVE`, rc 0 (`#include "vm.h"`; deleting -DEIGENSCRIPT_VERSION). The arm now replays each recorded call through the driver path resolved at startup: the recorded argv, from the recorded cwd, with its stdin, plus -fsyntax-only, minus the -o operand and the call's sibling TUs, each change argued in replay_argv_for_tu. The -m32 arm drops the same injected -Isrc/-D (same argument) and compiles from the recorded cwd. New self-test rows 10i/10d (real arm, against a wasm32 clang stand-in driver), 10im/10dm (-m32 arm), control 10c; each gut (re-add the injected flags to either arm) turns its rows red. Cases 57 -> 62. Gap 2 (feature): pages.yml's path-filtered PR trigger could never be a required check. It now runs on every PR, unfiltered, as ci.yml's pattern: `playground / scope` (the one path list), `playground / build` worker (same `bash web/build.sh`, steps guarded, a built=true receipt), and the aggregator `playground (real emcc wasm32 build)` (if: always(); a failed, cancelled or skipped worker, or a touched PR with no receipt, is failure). Deploy needs the aggregator and stays push-only. docs/CI.md names the aggregator as the check to require. werror_switch_check ilp32 floor 14 -> 13 (the real arm has no compile line of its own now; --print-counts said 13). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…ed?" decision Round 2's `playground / scope` decided "touched" from `git diff --name-only`, which under default rename detection lists only a renamed file's NEW path: moving src/trace.h -> attic/trace.h (or src/hash.c -> tools/hash.c) read as untouched, the build was skipped and the aggregator passed a recipe that no longer compiles (both critic families). Round 1's path filter was the same class. The build is ~72 s against a ~45 min CI run, so the decision is deleted, not patched: the scope job and its path list are gone, the worker runs `bash web/build.sh` on every PR and push with no condition upstream of it (only the Pages steps stay gated to non-PR events), and the aggregator `playground (real emcc wasm32 build)` is success iff the worker succeeded and left its built=true receipt. The aggregator is kept: a skipped job reads as passing to a required-check rule and a cancelled one leaves nothing; it makes both a failure and keeps the requirable name. Deploy needs it and stays push-only. docs/CI.md and CHANGELOG say what is now true; docs_claims PATHS|docs/CI.md 63 -> 60 (the deleted path-list paragraph). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
InauguralPhysicist
marked this pull request as ready for review
September 23, 2026 03:08
…ly one reason - tools/ilp32_syntax_check.sh header: the authoritative pages.yml build runs on EVERY pull request and push to main (round 3), not only on PRs touching src/web/docs/VERSION. - docs/CI.md: the check to require is the aggregator, "never the worker alone" — the `scope` job it named no longer exists. - Self-test plant 1: with -Isrc gone from the -m32 arm, the lone planted eigenscript.h copy could not resolve "value_slot.h", so the plant was red twice (the static assert AND a fatal missing include that masked every line past eigenscript.h:1666). The plant dir now stages src/'s headers (the recipe's own resolution, no -I re-injected), the UNPLANTED copy must compile clean there first, and the planted compile must yield exactly ONE error — the planted assert. Guts verified: staging only the lone header fails the control; bypassing the control too fails the one-error assertion (2 errors). Case count unchanged (62). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This was referenced Sep 23, 2026
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
1 participant
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.
Part 2 of #1255. The local gate [99i3] compiled with host
clang -m32(i386 ABI,doublealigned 4) while the playground is wasm32 (doublealigned 8): a layout_Static_assertcould pass the gate and fail the real build — measured 12 vs 16 bytes on a probe struct; the Value union was 36 vs 40. Following #1264 (build the real target, as mature languages do):web/build.sh) now runs on pull_request; Pages configure/upload/deploy stay main-only.tools/ilp32_syntax_check.sh: real emcc verdict when emcc is on PATH, otherwise an explicitly labelled APPROXIMATION pointing at the CI lane; new self-test rows 1w/1wt pin the i386-vs-wasm32 blind spot.Draft while the blind-critic round runs.
Refs #1255 #1264
🤖 Generated with Claude Code