The inspector learns which two boxes

[omars-lab]

inspector-draws-no-ink was filed as a question, not a defect: the encoding inspector shows identity and never appearance, and that blindness is what its authority rests on. The row said to wait for a caller rather than for an argument. The caller arrived — reading a 2→1 paint block at word 3, the next question is always which two boxes — so section ① now draws them.

What it draws, and what it still refuses to

Word boxes as outlines, from the committed assets/words/hafs-kfqc/<page>.json shards that gate:words already checks re-derive from committed bytes, on the page's own frame out of manifest.viewBox (with viewBoxOverrides honoured for pages 1–2). The rest of the page sits behind the ayah faintly, behind a toggle.

Nothing is read from assets/pages/**. No glyph is drawn. §6.4's principle keeps its force over the half that mattered, and the doc says which half: the line it actually protects is not "draws nothing" but "makes no claim about appearance", and a rectangle out of a gated shard makes none.

Three things settled before anything was drawn

• from === 1 is asserted, not assumed. True of all 6,236 ayahs — no ayah's boxes are split across two shards — which is what lets print word i be boxes[i - 1] unconditionally. A non-1 from would shift every label by the offset and look entirely correct.
• Boxes are tinted by host index, not print index. A word the fold dropped gaps on the page exactly as it gaps on the codepoint ruler. That shared parity is why ① and ④ read as one instrument rather than two pictures — and why clicking a box selects the annotation over it and lights the ruler in step.
• A count disagreement refuses to number. If the box list and the data-hafs word list disagree, the view draws the geometry and warns with both counts, because the failure it would otherwise produce is word 7's box under word 8's label. Measured over the full corpus: 6,236 ayahs outlined, 0 mismatches. The check stays; the number is a result, not a guarantee.

One invisible defect, found by building it

Both halves of the report are text to probe-encodings.mjs and are never imported, so nothing noticed a syntax error in the client: the report generated, weighed its usual 4.9 MB, and opened to a blank page with the whole script dead in the console. It now compiles the exact concatenation the browser will parse (new vm.Script) before writing, and fails at generation time.

Verified

Headless Chromium against the generated report: both themes render with zero console errors; page 2 uses the 235×235 override and page 7 the 345×550 default; clicking each of page 7's 17 boxes selects an annotation for exactly the 7 that carry one, with box 14 lighting 2 — the two-adjacent case the feature exists to answer; the context toggle drops rect.other from 132 to 0; injecting a fake extra box flips all 18 to .untrusted and surfaces the warning. make ci green.

Registers

④ goes to answered, not fixed — same reasoning as drift-label-reads-backwards. The guard catches a syntax error and gate:words keeps the shards honest, but nothing in CI would fail if ① stopped drawing: the report is generated, gitignored and never a gate, which is exactly why it is allowed to be this cheap.

§⑥ item 4 gets the superseded-blind-spot treatment rather than an edit in place, and map.json's extend rule inverts to where the line now sits, so the next person adding a font reads a rule that has already thought about their case.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EuhvbUKjGesE3uMhjCzBGt