epiphany/spec/CONTRACT_EDITOR_T4_SPIKE.md

44 KiB
Raw Blame History

Contract: Editor T4 — the toolkit spike

Revision 6 (2026-07-28), after five reviews. Revision 1 left discretion in the measurement and elimination mechanics; revision 2 reordered the ladder and pinned the deciding numbers, but its new structure carried its own defects — a damage oracle that compared the wrong thing, correctness failures dressed as environmental absences, and an escalation branch with no continuation. Revisions 3 and 4 closed those and introduced smaller ones of their own — a damage gate covering one rung, a censored median that flattered silence, and a latency threshold with no timer origin. Revision 5 closed those. Revision 6 amends Round 1 after building its oracle exposed two defects in the round as written: its glyph set was chosen by subpath count rather than by measured holes (fClef has none), and it framed criterion 1 around the fill rule when Bravura's oppositely-wound contours make even-odd and nonzero agree. Both are corrected below, and Ruling A criterion 1 is amended to match.

Repo root /home/jeans/Repos/active/epiphany. Governed by spec/PLAN_EDITOR_APP.md Ruling A ("What this ruling does not pin: the toolkit/tessellation stack. The T4 spike decides it, bounded by these recorded criteria" 16) and Ruling D (the app crate is created at T4, and the spike opens the tranche). All three §3.7 prerequisites are discharged: W1 dd33b34, W2 24f8c80, W3 f639919 (spec/ANALYSIS_TEXT_RUN_PRIMITIVES.md).

Execution model as T1a/T2/T3/W1/W2: Sonnet subagents implement per round, coordinator line-level review with independent re-runs, user deep-dives at this contract's sign-off, at every precommitted oracle before candidates run against it, at every measured table before it enters the ruling, and at the final verdict. Mutation discipline where there are assertions to mutate: anchor-assert before substituting, restore by reversing, never git checkout, never git stash.

Parallel safety. The genesis-operation tranche owns epiphany-core, epiphany-ops, epiphany-bundle (at G2), all .tex, and testkit's requirement-label constants, vectors, and generators (spec/PLAN_GENESIS_OPS.md). This packet touches none of them, and by pin 2 it modifies no existing crate at all. Requirement and conformance counts are reported as observed, never asserted from memory; this packet moves neither. Pin 12 additionally freezes the sources the spike measures against, which file-level non-overlap does not by itself achieve.


The verified starting point

Confirmed in the tree on 2026-07-28. Several of these make the spike smaller; two make it possible at all.

  • The tessellator's input already exists. W2's epiphany-glyphs exposes BravuraGlyphCatalog implementing GlyphCatalog with a real render_data, returning GlyphRenderData { outline: Vec<PathCommand>, .. } over PathCommand::{MoveTo, LineTo, CurveTo, Close} in staff-space units (glyphs/src/catalog.rs:21,39; layout-ir/src/glyph.rs:283). No candidate needs to parse SVG d strings, and no glyph work belongs in this packet.
  • Glyphs with holes are in the bundled set — but subpath count does not identify them. Revisions 15 named gClef/fClef/timeSig8/ accidentalFlat on the strength of "19 of the 37 bundled outlines have more than one subpath". That conflates multi-subpath with has a bounded hole, and fClef is the counter-example: its three subpaths are a bowl and two solid, disjoint dots, nested in nothing. Measured over all 37 by point-in-path (2026-07-28), exactly twelve carry a bounded holegClef, timeSig8, accidentalFlat, accidentalSharp, accidentalNatural, noteheadHalf, noteheadWhole, noteheadDoubleWhole, timeSig0, timeSig6, timeSig9, dynamicPiano — while fClef, cClef, barlineFinal and every repeat glyph carry none. Criterion 1 has real material, and Round 1 below names it by that measurement rather than by subpath count.
  • Bravura's contours are correctly oppositely wound, so even-odd and nonzero agree on every bundled hole. Signed ring areas, as measured by the Round 1 oracle's adaptive flattening (tolerance 0.0005 staff-space, the authoritative figures — magnitudes are flattening-dependent, the signs are the claim): gClef [8.702, 0.691, 1.803, 0.509], timeSig8 [2.674, 0.435, 0.515], accidentalFlat [1.040, 0.257], noteheadHalf [0.903, 0.368]; and fClef [2.534, 0.153, 0.148]all positive, which is the same fact from the other side: no counter, three filled components. The fill rule is therefore not the load-bearing property here; preserving every filled contour and every bounded counter is. Ruling A criterion 1 is amended to say so.
  • Per-system ownership is populated only by the real solver. Engraver's casting module builds one PrimitiveIndices per system (engrave/src/casting.rs:1185); the stub solver publishes everything unowned by deliberate honesty (layout-ir/src/resolved.rs:82-87). Damage-based repaint is therefore untestable under the stub — the spike uses Engraver, or it is not measuring Ruling A's mechanism.
  • ResolvedLayoutIR::systems() exists and is documented as adopted "at T4, not before" (resolved.rs). This packet is its first consumer.
  • There is no large fixture. The testkit's largest is two staves × twelve measures (two_staff_wrapping_pressure, fixtures.rs:346); the three-staff fixture is one measure (three_staff_close_content, fixtures.rs:225), and the ten_measure_* family is single-staff. No fixture is both wide and long. The spike builds its own ladder — pin 7.
  • There is no ResolvedText. W3 recommended disposition E and the .tex amendment is the core/layout track's work, unlanded. See pins 810.
  • Today's SVG exporter cannot serve as round 2's reference renderer. ShapedSegment carries font-internal glyph ids and "an ordinary SVG <text> element cannot request one" (ANALYSIS_TEXT_RUN_PRIMITIVES.md:496). W3 says the real exporter will need explicit-glyph path output. Pin 10 supplies that for the spike rather than letting the check be unrunnable.
  • The historical baseline is context, not a comparand. benches/editor_pipeline.rs (986c9cc, 4a4988c) measured today's SVG-string path at 123165 µs scene-build + 1.362.86 ms paint per edit — but dev profile, offscreen resvg, medians across score states whose ink differs by depth, and the bench documents these limits itself (editor_pipeline.rs:154). Pin 6 re-measures it inside the spike instead. The demo pins eframe = "0.29" and rasterizes through resvg into an egui texture (editor-gui/src/main.rs:361).
  • Two GPU adapters are present on the development machine — a discrete Navi 31 and an integrated Granite Ridge — with a live display. Vulkan 1.4 enumerated from an unsandboxed shell; re-confirm adapter enumeration from the spike's own process before relying on it, since a sandboxed run could not see it.
  • MSRV is 1.85, and the MSRV CI job excludes exactly one crate, epiphany-editor-gui (ci.yml:99). Pin 1 keeps that list unchanged.

What the spike is, and what it is not

It is an elimination experiment that produces one document: a ruling naming the toolkit/tessellation stack for epiphany-editor-app, with the measurements and disqualifications that chose it.

It is not the app. No command registry, no editing intents, no document layer, no undo, no goldens, no persistence. Ruling D creates epiphany-editor-app as a fresh crate after this verdict, built around the §3.5 command registry from its first commit; spike code is not grown into it, for the same reason the demo is not.

It does not reopen Ruling A. The architecture — direct vector canvas over ResolvedLayoutIR, viewport-culled, per-system damage, SVG demoted to export — is granted. The spike chooses the stack. A finding that no candidate can implement it escalates to the user as a ruling-amendment request; it does not silently promote the SVG path back.


Design pins

Placement and isolation

  1. The spike lives outside the workspace, at spikes/editor-toolkit/, as its own Cargo workspace with its own committed Cargo.lock, path-depending on the epiphany crates it reads. The root Cargo.toml gains one additive line — exclude = ["spikes"] — and nothing else. Rationale: candidates bring wgpu, vello, modern egui, lyon, rustybuzz; as workspace members they would enter the shared lockfile and the MSRV closure for code designed to be thrown away. (Cargo's exclude places the directory outside root --workspace and lets it keep its own lockfile.) Consequence, stated because it is a real cost: root cargo test --workspace does not build the spike, so the spike's own gate is run explicitly and reported.
  2. No existing crate is modified. The spike consumes public API only. If a candidate needs something layout-ir, glyphs, engrave, or editor-core does not expose, the spike works around it locally and reports the gap as a named finding for a later packet. An API that has to change to make a candidate work is itself evidence about that candidate.

Measurement

  1. Builds are --release --locked. Every timed figure, without exception. A dev-profile number decides nothing here.
  2. One common deciding configuration, identical across candidates: offscreen render target (not a surface), fixed target size 1920 × 1080, identical MSAA sample count (4×, or the highest all survivors support, stated), identical sRGB target format, wgpu Vulkan backend on both adapters. Candidates are compared only within this configuration. Surface-present timing is measured separately as a capability figure and decides nothing — it mixes in the compositor.
  3. Timed work must be completed GPU work, not submission. Each timed iteration ends with an explicit device wait (Device::poll with Maintain::Wait, or the candidate stack's equivalent barrier) and, where the backend supports timestamp queries, a GPU-side timestamp span reported alongside the wall figure. A candidate that cannot be made to report completed work is reported as NOT RUN for that figure, never as fast. Protocol per figure: ≥50 warm-up iterations discarded, then ≥200 timed iterations; report median, p95, min, max, and iteration count. The deciding statistic is the median, with a predeclared practical equivalence band of 10% — a band chosen in advance to stop small differences deciding architecture, not a claim of statistical indistinguishability, which this protocol does not test for. Vsync is off, or the figure is a full distribution — a 16.7 ms mean under vsync measures the display.
  4. The SVG comparison is re-measured inside the spike, or it is not made. Under pins 35, on the exact same fixture, target, size, statistic, and machine: render the score to an SVG string, parse with usvg, rasterize with resvg to the same offscreen dimensions, and compare scene-build + paint totals, because the premise under test concerns the whole renderer-owned path, not paint alone. The historical 123165 µs / 1.362.86 ms figures appear in the report labeled as dev-profile historical context only and are never compared against a candidate. What falsification means, defined in advance: the SVG path has no damage mechanism — it re-serializes and re-rasterizes everything per edit — so the honest comparison is candidate damage repaint against SVG full repaint, at the deciding rung. A candidate losing on full paint at the smallest rung does not falsify Ruling A's demotion, since damage and culling are exactly what the demotion buys; a candidate losing on the damage-vs-full comparison at the deciding rung does, and escalates to the user as a finding rather than resolving itself.
  5. The fixture ladder is pinned now, before any candidate renders, so the deciding workload cannot move. Rungs are F1F4 (named apart from the elimination rounds, which are 05): F1 1 staff × 10 measures (bench parity), F2 4 staves × 32 measures, F3 12 staves × 100 measures, F4 24 staves × 200 measures. Generated in the spike workspace (not testkit, whose generators the parallel track owns) and engraved up front with Engraver::default(). Dimensions do not pin the workload, and the workload is what gets tessellated. The generator's musical-content recipe — notes per measure, accidental cadence, slur and tie density, voice count, articulations, and the fixed seed — is committed to the spike repo and user-reviewed before any candidate-specific implementation begins, under the same rule as the oracles (pin 13). One recipe, scaled across rungs by dimension only. Reporting: per rung, actual page and system counts, per-type primitive counts (glyphs / strokes / curves) and total path-command count — not an aggregate primitive number. Path commands are what a tessellator actually consumes; two layouts with equal primitive counts can differ severalfold in curve work. Rung validity and the deciding rung are decided independently of any candidate. A rung is dropped only by the shared engraving preflight — engrave exceeding 10 minutes wall, or memory exhaustion — reported with its reason. The deciding rung is the largest engraving-valid rung, full stop. A candidate that cannot render there records its own NOT RUN and the eligibility consequence in pin 16; one candidate's inability does not shrink the contest for the others, which revision 2's "largest completed by every survivor" wording allowed.

Text

  1. Text uses a local, explicitly non-canonical stand-in. SpikeResolvedText in the spike workspace, mirroring W3 §3E completely: provenance; source string; shaping identity (pin 9); Vec<ShapedSegment> of positioned glyphs — each with face index, source range, direction, script, language, size, and per-glyph transform; cluster map; measured bounds; reserved_box; origin; align; style; layer. Mirroring a subset and calling it §3E would test a shape the amendment is not going to have. It is not the .tex amendment and does not pre-empt it — the reverse is the point. The spike is that shape's first consumer, so every place it proves awkward to consume is a finding routed back to the amendment, which is the cheapest available review of a design otherwise reviewed only on paper.
  2. The stand-in identity records every field of W3's TextShapingIdentity (ANALYSIS_TEXT_RUN_PRIMITIVES.md:407), not a file hash: the ordered fallback chain, and per face its family and version (W3 carries both as diagnostics even though the content hash is the identity that binds), its file content hash, face index, variations, and synthesis flags; the shaper implementation id and version ("it moves glyphs, so it is an input on exactly the footing of the font version"); the OpenType feature set applied, in canonical order; and the Unicode version governing both the bidi algorithm and grapheme segmentation. The exact fixture strings are committed verbatim in the spike repo. A partial identity would let two runs agree on pixels and disagree on clusters — the divergence W3 §5 check 4 exists to catch. Shaping is fixture generation, not a recommendation: rustybuzz shapes and unicode-bidi itemizes. unicode-bidi does not do grapheme segmentation — the segmentation implementation is separate (unicode-segmentation or equivalent) and it, and its Unicode-data version, are named in the identity and the report, because caret stops come from it and not from the shaper. Whether the engraver adopts any of these is the core/layout track's call; the spike reports on fitness only. Faces resolve once at startup from an explicit path list, with their bytes hashed — W3's narrowed rule exactly ("a host face may participate only once resolved to an exact content-hashed asset"), committing no font binary. A required face absent on the machine ⇒ NOT RUN (pin 14).
  3. The spike builds its own SVG reference emitter. Today's exporter cannot draw a SpikeResolvedText<text> cannot request font-internal glyph ids (ANALYSIS_TEXT_RUN_PRIMITIVES.md:496) — so without this, round 2 check 1 would be NOT RUN for every candidate and the round would decide nothing. The spike emits explicit glyph outlines as <path> from the same hashed face and the same glyph ids (ttf-parser, already in rustybuzz's tree), rasterized with resvg under pin 4's configuration. This is a prototype of the explicit-glyph output W3 says the real exporter needs, and its findings are reported as such. Rejected alternative: sequencing round 2 after the real exporter lands — that would block T4 on the core track's amendment, which is precisely what pin 8 exists to avoid.

Damage

  1. Damage repaint is defined as a state transition, with an oracle. For the fixture at each rung, build a deterministic A/B pair: score S and S, where S is S with one operation applied to a single note chosen to lie inside one system. Engrave both up front. The damage set is computed from rendered content, never from raw index vectors. PrimitiveIndices are positions in flat arrays (resolved.rs:77), so comparing the vectors is doubly wrong: a pitch change that preserves primitive count alters geometry while every index vector stays identical, and an inserted accidental renumbers global indices so that later systems' vectors differ although nothing they draw moved. Instead, each system is fingerprinted over its dereferenced primitives — glyph reference, quantized position, transform, bounding box, style, layer for glyphs, and the analogous fields for strokes and curves — together with the system's own geometry, under a fingerprint function committed in the spike repo. Assert before timing that exactly one system's fingerprint differs between IR_A and IR_B. If the count is anything other than one — zero included — advance to the next note in a finite, committed candidate order (notes in canonical layout order, first 20 tried) and record how many were tried. Zero matters as much as many: an operation whose layout effect the solver absorbs entirely would otherwise yield a "damage repaint" that repaints nothing and times as instant. Exhausting the candidate order is a harness/preflight failure — reported against the spike, not against any candidate, and not recorded as candidate evidence, since every candidate shares the same A/B pair. unowned is fingerprinted the same way, not merely counted — size alone misses changed geometry at unchanged indices. If its fingerprint differs, its primitives join the damage set and the report says so; its size is reported at every rung regardless, because a large unowned bucket caps what damage repaint can ever save, and hiding that would flatter every candidate equally but the architecture not at all. Timed: retessellation of the damaged system + cache replacement + completed paint (pin 5), starting from a warm scene built from IR_A. Oracle: the damage-updated target must be byte-identical to a from-scratch full paint of IR_B in the same process and configuration. (Byte equality is legitimate here — same renderer, same settings — unlike the cross-renderer comparison in pin 6, which is a bounded visual differential.) Failing the oracle is a FAIL, not a NOT RUN. Under pin 14's split: the damage capability cell is FAIL — the check ran and the output was wrong — while the damage timing is NOT RUN, because no valid timing exists for an incorrect result. Revision 2 recorded both as NOT RUN, which filed a broken implementation as an environmental absence. A damage-capability FAIL escalates with a required root-cause attribution: candidate limitation, or spike implementation defect. That distinction decides whether it eliminates — Ruling A's granted architecture is per-system damage, so a genuine candidate limitation is disqualifying, while a spike defect is a bug to fix and re-run — and it is not a judgement the report author makes silently.

Inputs and outcomes

  1. The measured source tree is frozen and named. All rounds run inside a dedicated git worktree checked out at a pinned baseline commit, recorded by SHA in the report, so the spike's path dependencies resolve to frozen sources rather than the working tree — which currently carries the parallel track's in-flight edits. The spike's own tree is placed into that worktree (its ../../crates/... path deps then resolve inside it), so the only unfrozen inputs are the spike's own files. Builds are --locked. If the baseline moves, the affected round is re-run in full; partially re-run rounds are not reported. Blast radius's "only modification to pre-existing workspace configuration or production files" is scoped against that baseline, not against the globally dirty tree.
  2. Every check has a precommitted oracle. Oracles are committed to the spike repo, with their commit recorded, before any candidate renders against them. This is a hard sequencing rule, not a preference: "points that must be ink" chosen after seeing output is not a test.
  3. The outcome model, in three parts. (a) Capability results are PASS / FAIL / NOT RUN. NOT RUN means could not execute — missing font, missing backend, unavailable API, environment absent. A check that ran and produced the wrong answer is FAIL, never NOT RUN; conflating them lets a broken implementation buy an escalation instead of recording negative evidence. (b) Timings are measured or NOT RUN, separately from (a). A timing is NOT RUN whenever no valid figure exists — including when the corresponding capability FAILed, since timing a wrong result measures nothing. (c) Eligibility is tracked as "disqualifying checks passed", separately from criterion cells. The disqualifying set is fixed here: round 0's accessibility route; round 1's fill correctness; round 2 checks 2 and 5; round 3's semantics; and a damage-capability FAIL attributed to candidate limitation (pin 11). A criterion cell is the worst of its checks and is reported for the record — but eligibility for the tie-break is the disqualifying set alone, so a FAIL on a non-disqualifying check does not silently strand a candidate the user chose to keep, which revision 2's "passing every hard criterion" wording did. Criterion 2 has no PASS/FAIL cell at all — it is a ranking criterion, represented in the matrix by its measured table plus the damage-capability cell, and a fabricated pass for it would be a fake row. The spike is not complete while a disqualifying check is NOT RUN for a surviving candidate, and a NOT RUN that would decide the verdict escalates (pin 16) — with one exemption, named here so this pin and pin 16 cannot drift apart: a NOT RUN timing at the deciding rung is the predeclared ranking loss of pin 16 step 1 and eliminates without pausing.
  4. Current releases, not Ruling A's snapshot. Ruling A criterion 6 records "egui 0.29 / 0.35" as a 2026-07-23 observation. The spike evaluates each candidate's actual release at spike time and records version + release date as a matrix row. Criterion 6 is a question the ruling answers, not a test a candidate passes — Ruling A defers it "to the spike's call under criteria 15" — so it gets no PASS/FAIL cell and is answered in prose, with the version row as its evidence.
  5. The tie-break, fixed before measurement, with escalation as a real outcome. Among candidates whose disqualifying set (pin 14c) is fully passed, in order:
    1. Damage-repaint median at the deciding rung (pin 7), integrated adapter, pin 4's configuration. Anchored to the best median, not pairwise: retain every candidate within the 10% band of the best, discard the rest. Pairwise comparison is non-transitive with three candidates and could order them inconsistently. One survivor ⇒ it wins. A candidate with NOT RUN at the deciding rung is not retained here and its inability is reported as such (pin 7). This is a predeclared ranking loss and is explicitly exempt from the escalation rule below — the deciding rung is the largest engraving-valid rung, so failing to render it is failing on the merits at the scale the product targets, decided by a rule fixed before any measurement. If step 1 retains nobody — every eligible candidate NOT RUN at the deciding rung, so there is no best median to anchor to — that is the no-winner branch, not an empty selection: the spike stops and the ruling records a ranking-loss wipeout with the recommended widening.
    2. Foreclosure coverage — count of round-5 probes passed (03), each binary. Retain every candidate at the maximum count and discard the rest; one survivor ⇒ it wins, multiple survivors ⇒ go to 3. (Stated because "equal counts ⇒ go to 3" left 3/3/2 undefined.)
    3. Maintenance surface, by Pareto rule over three operationally defined axes. The measured upstreams are pinned here, before any figure is fetched, because every candidate is a composite and "the primary crate" would let the favourable upstream be chosen after seeing the data: C1 = egui (emilk/egui) + lyon (nical/lyon); C2 = vello (linebender/vello) + winit (rust-windowing/winit); C3 = iced (iced-rs/iced). Where a candidate names more than one upstream, it takes the worse value on each axis — conservative, and it removes the selection entirely. Each figure is recorded with its measurement date:
      • Transitive crate count — unique packages from cargo tree --edges normal --target <host triple> for the candidate's spike crate under exactly the features the spike enables, excluding the epiphany path crates. Lower is better.
      • Release recency — days since the most recent crates.io release that is neither yanked nor a prerelease. Lower is better.
      • Issue responsiveness — measured on the pinned repositories above, at a recorded snapshot date, from an archived issue/comment snapshot committed to the spike repo so the figure is recomputable after the tracker moves on. A qualifying issue is one opened in the 180 days before the snapshot, in the issue tracker (pull requests excluded), not authored by a maintainer or a bot. Maintainer is machine-observable, not inferred: a comment whose recorded GitHub author_association is OWNER, MEMBER, or COLLABORATOR, or whose author is listed in a named owners file at a pinned commit. Triage/write permission is not reliably public and is not used. Bot comments never count as a response. Take the twenty most recent qualifying issues and the time from open to first maintainer response. An issue with no maintainer response contributes +∞, not its current age. Revision 4 used opening-to-snapshot age, which is only a lower bound on the response time and would have given twenty issues opened yesterday and ignored an excellent median. With +∞, the median stays computable while fewer than half are unanswered and becomes +∞ — the worst possible value, which is the right answer — once half or more are. Lower median is better; fewer than five qualifying issues ⇒ this axis is NOT RUN and step 3 escalates. A candidate wins step 3 only if it Pareto-dominates every other remaining candidate — no worse on all three axes and better on at least one, against each of them individually. No aggregate score, no weighting, no author discretion. A decisive NOT RUN, a tie surviving all three steps, or any Pareto-incomparable pair escalates rather than selectsexcept the deciding-rung NOT RUN at step 1, which is the predeclared ranking loss above and eliminates without pausing.

The candidate set

Three entrants — all three named by Ruling A criterion 6 (PLAN_EDITOR_APP.md:596: "modern egui, iced, or a Vello surface"). The spike may add a fourth with written justification in the report; it may not drop one silently. Exact crate names, versions, features, and backends are recorded in the report and frozen by the committed Cargo.lock.

  • C1 — modern egui + lyon-tessellated meshes. Closest to the demo, so the migration story is cheapest; the open questions are whether pushing tessellated meshes through epaint holds up at F3/F4, and whether egui's text stack can be bypassed cleanly for pin 8's pre-shaped runs.
  • C2 — vello behind a winit shell. GPU-compute path rendering, so criterion 1 is likely free; the open questions are windowing and UI chrome (there is no widget toolkit), accessibility wiring, and maturity.
  • C3 — iced. Retained-mode where C1 is immediate-mode, which bears directly on per-system damage, and it ships its own wgpu renderer. Accessibility is its live risk, and round 0 must prove the route rather than assume it: AccessKit lists egui among integrated projects and does not list iced, and iced's upstream accessibility issue (#552) remains open. A manual accesskit_winit route may exist; round 0 demonstrates it or C3 fails there.

Named exclusions, so their absence is a decision. skia-safe: a C++ build chain against a project whose MSRV and reproducibility posture rests on a pure-Rust closure, plus a vendoring and licensing surface out of proportion to the gain. GTK/Qt drawing surfaces: excluded on deployment, ABI stability, dependency weight, and integration costnot on any claim that they foreclose overlays, freehand input, or touch, which they do not. Either may be reconsidered if all three entrants fail.


The elimination ladder

Cheapest hard disqualifier first, genuinely: every disqualifying check (pin 14c) is settled before the expensive ranking round runs. All candidates complete round n before any enters round n+1.

Round 0 — accessibility route + desk survey (HARD). For each candidate record: current release and date, transitive dependency count, MSRV, and whether the fill tessellator documents nonzero/even-odd support. Then the hard part: demonstrate an accessibility route — a minimal window exposing one accessible node whose role and name are read back through the platform adapter, for every candidate without exception. Naming an integration crate does not satisfy this round; it is at most a hint about how hard the demonstration will be, and revision 2's wording that let a crate name substitute for a readback was a false-pass path. A first-party integration and a manual accesskit_winit wiring are equally acceptable routes; only the readback is the evidence. Timebox: two working days per candidate, fixed here rather than left open — exceeding it is a FAIL with the attempt and its blocking point documented, not an indefinite extension. Rationale: criterion 4 is a hard criterion, and discovering its absence in round 3 would waste three rounds — which is the whole claim of "cheapest disqualifier first".

Round 1 — criterion 1, compound-path fill correctness (HARD). Each candidate draws five glyphs from BravuraGlyphCatalog's typed outlines under pin 4's configuration, in two check classes testing two different properties. All sample points are derived programmatically by point-in-path over the PathCommand outline — never chosen by eye — and every point must lie ≥8 device pixels from any outline edge at the pinned render transform, so antialiasing cannot explain any result.

  • Hole checks — gClef, timeSig8, accidentalFlat, noteheadHalf. ≥3 must-be-ink and ≥3 must-be-background points, each inside a bounded hole — enclosed by the glyph's outer silhouette yet unfilled — asserted by the derivation, not assumed. A background point merely outside the silhouette tests nothing, since a renderer that fills holes solid passes it trivially. noteheadHalf is in this set deliberately: it is a frequently repeated and semantically consequential glyph — a filled counter renders half notes as quarter notes, a notation error rather than a cosmetic artifact. (It is not the single most-drawn glyph; noteheadBlack is, and it has no counter.)
  • Disjoint-component check — fClef. It has no bounded hole (bowl plus two solid dots), so it carries no background requirement and tests the other half of compound-path correctness: ≥1 ink point inside each of its three filled subpaths — body and both dots — at the same clearance floor, each tagged with its subpath_index so the oracle proves coverage of every component rather than three generic ink points that could all land in the bowl. A tessellator that keeps only the largest contour fails here and would pass every hole check.

The oracle's status model must be explicit, not inferred from an absence. Each glyph carries a requirement enum (or at minimum background_required: bool) plus an overall satisfied status. fClef passing with zero background points is a satisfied result under its own requirement class; recording it only as background_satisfied = false would make a correct outcome indistinguishable from a failed one.

The oracle file, the render transform, and the expected class per point are committed before any candidate renders. Fails hard: any bounded hole painted as ink, or any required filled subpath omitted.

Round 2 — criterion 3, text (HARD). W3 §5's five checks against pin 8's stand-in and pin 10's reference emitter: (1) faithful consumption — draws positioned segments without re-shaping, matching the reference emitter's rendering of the same run under Ruling A's bounded visual differential; (2) fallback, forced — resolves through the declared chain only, and reports rather than substitutes an uncovered codepoint; (3) bidi — a mixed Arabic/Latin run itemizes into multiple directional segments, each drawn in its resolved face at its resolved position; (4) hit testing at character granularity — UTF-8 byte offsets as base index, caret stops at grapheme-cluster boundaries with bidi affinity; (5) accessibility — the run appears in the tree as its source string, not a graphic. Outcome rule (pin 14): the criterion cell is the worst of the five. Checks 2 and 5 are in the disqualifying set — W3 names those two as disqualifying "regardless of tessellation throughput" — so failing either eliminates. Failing 1, 3, or 4 marks the cell FAIL and is not disqualifying by default, since W3 did not make them so and this contract does not silently promote them. What that escalation is, concretely: the spike pauses and the user issues a recorded waiver or ruling amendment that states explicitly whether the failed check joins the disqualifying set for this candidate. Eligibility then follows pin 14(c) from that record. Revision 2 left the user a decision with no effect — the candidate stayed FAIL on the criterion and pin 16 admitted only candidates passing every hard criterion, so a "keep it" ruling could not reach the tie-break.

Round 3 — criterion 4, accessibility semantics (HARD). One score fragment exposed with meaningful semantics. Precommitted oracle (pin 13): a fixed numbered interaction script — key and pointer actions in order — and for each step the required evidence: accessibility-tree node role, name, state, and supported actions, plus before/after focus and selection state, plus the transcribed spoken output. Orca version, AT-SPI version, and desktop session are recorded. Transcription alone does not satisfy this round; navigation, focus movement, selection updates, and command activation must each be evidenced by tree state, because a stack can narrate while remaining unusable. Fails hard: a tree that exists but conveys nothing — Ruling A names that outcome as failure, so carrying AccessKit earns no credit.

Round 3b — damage correctness (HARD, untimed). Pin 14(c) makes a candidate-attributed damage failure disqualifying, so it must be settled before the ranking round, not discovered inside it — otherwise the ladder's "every disqualifier before round 4" claim is false. Each survivor runs pin 11's A/B transition at every engraving-valid rung with the byte-identity oracle, and nothing is timed: this round asks only whether correct per-system damage is implementable in the stack. Every rung, not F1 alone — revision 4 checked F1 and left an F2F4 oracle failure to surface inside the ranking round, which is exactly the late hard disqualification this round exists to prevent, and larger rungs are where multi-system and unowned handling actually breaks. Byte equality is necessary but not sufficient, so it is paired with a reuse assertion: the candidate's scene cache is instrumented to record which system representations were rebuilt during the transition, and the rebuilt set must equal the damage set. Without it a candidate that silently redraws everything produces a byte-identical target and passes as if it implemented damage — it would then be caught only by round 4's timings, i.e. by being slow, which is evidence about performance rather than about architecture. Outcome is the damage-capability cell (pin 14a), with root-cause attribution on failure. Still cheap: no measurement protocol applies, so each rung is one transition rather than 250 timed iterations. Round 4 then times damage for candidates that already proved it correct.

Round 4 — criterion 2, staged latency (RANKING, no hard pass). Survivors only. Per rung (pin 7), per adapter, under pins 35: (a) scene build — IR to the candidate's draw representation; (b) full paint; (c) damage repaint per pin 11, with its oracle; (d) viewport-culled paint at 1920 × 1080 on the deciding rung — with the viewport transform and location committed in advance alongside the oracles (pin 13) and the visible primitive count reported, since a size without a position leaves the culled workload unpinned. Plus the SVG re-measurement of pin 6. The integrated adapter's figures decide; the discrete adapter's are reported as headroom — a budget met on a Navi 31 says nothing about the machine most users have. Every candidate's figures within a round come from the same spike commit, recorded in the report, in addition to pin 12's frozen root baseline and the oracle commits — otherwise the harness itself is an uncontrolled variable between candidates.

Round 5 — criterion 5, foreclosure probes (RANKING). Prototyped, not assumed, each with a binary pass condition fixed here: overlay — a moving presence cursor drawn in a separate layer while the score's damage set stays empty (asserted, not observed); freehand — a live stroke rendered from a committed synthetic pointer trace (≥1,000 events, replayed at a stated fixed rate, committed with the oracles under pin 13), on the integrated adapter at the deciding rung under pin 4's configuration. Latency is measured from the trace's scheduled injection timestamp to completion of the first frame containing that event — backlog included. Starting the timer when the renderer picks the event up would let a candidate buffer the whole trace, render it late, drop nothing, and report fast update work while lagging visibly; that is the failure mode this probe exists to catch, so the timer origin is the schedule, not the pickup. Coalescing is permitted and must be declared: where a frame covers several injected events, each of those events is charged that frame's completion time. Every sample in the trace must appear in the resulting stroke geometry — asserted against the committed trace, not assumed. Passes if the p95 of that latency is ≤ 16.7 ms — the spec's own frame figure, not an invented one — and no injected event is dropped or missing from the stroke; a single one fails the probe. (Revision 3 pointed at "round 4(c)'s budget", but 4(c) defines a measurement, not a threshold, and this probe feeds the deciding tie-break.) touch — multi-touch events received with distinguishable pointer ids, and a two-finger pan/zoom driven by them. Coverage is the count passed (03) and feeds tie-break step 2 only.


Blast radius

New spikes/editor-toolkit/** (its own workspace, committed Cargo.lock, candidate sub-crates, fixture generator, SVG reference emitter, precommitted oracle files, and a DECISIONS.md); one additive exclude = ["spikes"] line in the root Cargo.toml; spec/RULING_EDITOR_TOOLKIT.md and this contract. Nothing else — no epiphany-core, no epiphany-ops, no epiphany-bundle, no epiphany-layout-ir, no epiphany-glyphs, no epiphany-engrave, no epiphany-editor-core, no epiphany-editor-gui, no testkit, no .tex, no CI change, no new workspace member, no golden touched or blessed. Scoped against pin 12's baseline commit.

Gate (actual output)

The root workspace must be unchanged in behavior: cargo fmt --check, cargo clippy --workspace --all-targets -- -D warnings, cargo test --workspace, conformance suite and requirement counts reported as observed and unmoved, all GUI goldens byte-identical, layout canonical bytes byte-identical. The one-line exclude is the only modification to pre-existing workspace configuration or production files against the baseline — the contract, the ruling, and the spike tree are additions at the repository root too, so "the only root diff" would have been wrong — and the gate proves it changed nothing.

The spike workspace is gated separately and explicitly, since root --workspace does not reach it: cargo fmt --check and cargo test --locked inside spikes/editor-toolkit/, reported with its own output. Clippy is advisory there, not -D warnings: throwaway code held to the production bar wastes the reviewer's attention, and pin 1 guarantees none of it ships.

Deliverable

spec/RULING_EDITOR_TOOLKIT.md — the verdict. A candidate × check matrix with every capability result carrying PASS / FAIL / NOT RUN and its evidence, its timing carrying measured / NOT RUN separately (pin 14ab), criterion cells derived from it, an explicit disqualifying-set column recording eligibility (pin 14c), no PASS/FAIL cell for criterion 2, and a version row answering Ruling A criterion 6 in prose (pin 15); round 4's tables in full with adapter, configuration, rung, per-type primitive and path-command counts, statistic, and iteration counts on every figure; the pin-12 baseline SHA, the spike commit per round, and the oracle commits; the tie-break applied on the record per pin 16, including any waiver or amendment that changed eligibility; the named foreclosures the choice accepts; and round 2's findings routed back to the ResolvedText amendment. Plus the no-winner branch if it fires, in either of its two forms: disqualification — which disqualifying check killed which candidate — or ranking-loss wipeout, where no candidate rendered the deciding rung so step 1 retained nobody (pin 16). Both end in a recommended widening; the second is not a lesser outcome and is not reported as an inconclusive run.

Report

Per round, as every tranche: files + summary, exact measured values with their full labels, every NOT RUN with the reason it could not run, gate output for both workspaces, deviations flagged explicitly. The user reviews each precommitted oracle before candidates run against it, and each measured table before it enters the ruling.