engine: Position is an integer cell — rounds and clamps to i16 by Morgandri1 · Pull Request #22 · Morgandri1/wheel

Morgandri1 · 2026-09-06T18:01:08Z

Summary

Implements the operator ruling in ARCHITECTURE.md ("Position is an integer cell", 2026-09-06): Position{x,y} changes from f64 to i16, with a custom Deserialize that accepts any JSON number (an in-flight drag can still send a float), rounds to the nearest cell, and clamps to i16::MIN..=i16::MAX before the value is ever stored, compared, or serialized back out.
Regenerated docs/schema/*.json (position.json's field type is now integer/int16; additionalProperties: false preserved via #[serde(deny_unknown_fields)] on the struct even though Deserialize is hand-written).
Updated the two contract-pinning tests that asserted the old float shape (node.rs, json_shape.rs) and added four new tests for round/clamp/JSON-parse/deny-unknown-fields, mutation-checked (reverted clamp_cell to a bare cast, confirmed the round/clamp tests fail).

Why

qa/integration/test_node_position.py's POS-is-an-integer/POS-rounds-and-clamps cases are currently failing on main's own tip — confirmed on a fresh CI run of main (not caused by any of my other open PRs). wheel-core::Position had no rounding/clamping logic anywhere; only the web-side UI clamp had landed.

Test plan

cargo test -p wheel-core (55 lib + 63 integration tests, including the new Position tests) — green
cargo test -p wheel-engine -p wheel-cli -p wheel-api — green, no code changes needed outside wheel-core (Position flows through Position::new(f64, f64) unchanged at every call site)
cargo build --workspace — clean
cargo fmt --check -p wheel-core, cargo clippy -p wheel-core --lib --tests — clean
CI: qa/integration/test_node_position.py POS-* cases should go green

vercel · 2026-09-06T18:01:11Z

Deployment failed for project wheel with the following error:

Resource is limited - try again in 24 hours (more than 100, code: "api-deployments-free-per-day").

Learn More: https://vercel.com/morgandri1s-projects?upgradeToPro=build-rate-limit

Rust's f64::round is half-away-from-zero; JS Math.round is half-up. They agree everywhere except a negative exact half: Math.round(-10.5) is -10, and the engine will store -11. Board coordinates go negative the moment you pan, so that is reachable, and it is one cell of permanent disagreement on every dragged node — invisible on screen, and exactly the drift the position gate is meant to catch. Fixed before #22 lands rather than after, so the engine's arithmetic and ours are written from the same rule instead of converging on a bug report. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Hw5v8WJBYWzZ1SYaD5xctn

Rounding cannot -- 0.5 cells/axis, one cell is one CSS px, canvas.tsx caps zoom at 1.8, so 1.27 px worst case. That is geometry and holds for any board. Clamping is unbounded and is a question about the DATA, not the code, so it can only be answered by looking at the rows that exist. Takes a sqlite path or a live engine URL, names any node outside the i16 bound and prints how far it will jump in pixels. Verified in both directions: exit 0 on a clean board, exit 1 naming a runaway node. Written because I asked PM for a SQL query and a request someone has to translate is a request that waits. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01F1nLndrG8rrdr5ar7P6WkN

… way in Implements the operator ruling in ARCHITECTURE.md (position cell size and clamping, 2026-09-06): Position{x,y} is now i16, with a custom Deserialize that accepts any JSON number, rounds to the nearest cell, and clamps to i16::MIN..=i16::MAX before a value is ever stored or compared. Fixes the POS-is-an-integer/POS-rounds-and-clamps failures currently red on main (qa/integration/test_node_position.py).

Position's exported wire type became integer/int16 in the previous commit (the operator ruling). JSON Schema's type: integer keyword correctly rejects a non-whole number like -1234.5 by spec, even though the engine's Deserialize impl still accepts and rounds it. That's an engine-level behavior the static schema check can't express, so this fixture is marked _pending the same way tool_source_postman.json already is, pointing at the wheel-core unit test that actually covers rounding.

PM approved landing this: additive to QA's docker-based ENG-starts- without-shm gate (qa/integration/test_engine_shm.py), not a duplicate -- same failure mode, different technique, no docker required. QA's own comment says the directory-at-the-shm-path fixture "verified to reproduce the production symptom exactly," and that fixture is a filesystem fact, not a container one, so it reproduces here too. Seeds a database already in WAL (the shape that actually crash-looped production -- journal_mode persists in the header, so every boot rediscovers it), blocks the -shm path with a directory, then asserts open_configured(path, false) -- the engine's own mode -- both returns Ok AND accepts a real write, with the pre-existing row intact. Mutation- checked: reverting mode_holds to a bare read-back (dropping the `BEGIN IMMEDIATE; COMMIT` proof) reproduces the exact symptom -- settles back on "wal", the mode that just failed. Does not close qa/BUGS.md #22 -- that stays QA's call once they confirm 1163cc9 live against a fresh image (PM is asking them directly).

…raction Content review of the rescued tree, per PM. All four pieces were GENUINELY NEW — none of this text is on main — and the classification is per piece rather than per commit, because the histories diverged and replaying them would have carried a claim its own author withdrew. CODE (still needed, verified against main): coverage_gate printed `r.stderr[-2000:] or r.stdout[-2000:]` on failure — stderr only when stderr is non-empty, and only its last 2000 chars. cargo puts the failing test's name and panic in the MIDDLE of a long stream, so the slice reliably cut the one thing worth reading. I hit this myself tonight: I could not read a coverage failure from CI and dug through the whole log twice. Main still had the bad slice; it now finds where the failure starts. BUG-022 CLOSED, by two people who each had half the proof. That session proved the mechanism at the crate level (mode_holds before the fast path, directory blocking -shm) and wrote down honestly that it could NOT verify through the docker fixture, having no docker binary. That gap is closed from the other side: ENG-starts-without-shm is green through the fixture, engine-shm 12/12. Neither session could have produced both halves. BUG-026 CLOSED by measurement — POS-* is 34/34 since #22 landed. Main's entry still said "17 red", true when filed and not any more. BUG-034, THE COLLISION, resolved by renumbering: that session filed the table-columns finding as 026; main had already given 026 to the position ruling. Different bugs, same number. Renumbering is the whole fix. AND THE PART THAT WAS ACTUALLY IRREPLACEABLE — the uncommitted file. Commit c9f0f6f claims the columns bug CONFIRMED, citing a CI run. The uncommitted correction RETRACTS it: PR #12 predates PR #20 and tested a route that 404s unconditionally, so neither run was evidence about columns at all. Landing the commits without the uncommitted file would have put a CONFIRMED claim on main that its own author had already withdrawn, and sent the next person hunting a bug the evidence never supported. All three states are recorded — claimed, retracted, then settled green by measurement. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01F1nLndrG8rrdr5ar7P6WkN

…elete Content review complete. All four pieces were genuinely new; none of the text was on main. Classified per piece rather than per commit, because replaying the commits alone would have landed a CONFIRMED claim its own author had already retracted in the uncommitted file. coverage_gate the blind stderr tail-slice that hid the failing test's name BUG-022 CLOSED — crate-level proof from that session, fixture-level proof from this one; neither had both halves BUG-026 CLOSED by measurement — POS-* 34/34 since #22 BUG-034 the 026 collision, renumbered; claim, retraction, and the measurement that finally settled it Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01F1nLndrG8rrdr5ar7P6WkN

Morgandri1 force-pushed the sdk/position-integer-clamp branch from 052e08f to 303cfdb Compare September 6, 2026 18:12

SDK/Engine added 2 commits September 7, 2026 00:20

Morgandri1 force-pushed the sdk/position-integer-clamp branch from 303cfdb to 7b411ee Compare September 7, 2026 00:21

Navigation Menu

engine: Position is an integer cell — rounds and clamps to i16 - #22

engine: Position is an integer cell — rounds and clamps to i16#22
Morgandri1 wants to merge 2 commits into
mainfrom
sdk/position-integer-clamp

Morgandri1 commented Sep 6, 2026

Uh oh!

vercel Bot commented Sep 6, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

Morgandri1 commented Sep 6, 2026

Summary

Why

Test plan

Uh oh!

vercel Bot commented Sep 6, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant