[robfeldmann]

Slice 6 of plan 036 (docs/plans/phase-4-view-layer-and-developer-tooling.md): the projection, the SEL locators, the invariant rules, expect cost, the trace, shot as text plus grid, and the first per-primitive demos. Stacked on #66 (phase-4/05-scenarios), so this PR targets that branch.

What lands

• Projection (Sources/TesseraHarness/Projection): the view graph's diagnostics as sorted-key JSON with a version, plus the screen as rows and a style grid. dump prints it.
• Rules v0 (ids are the 1.0 promise): no-dirty-after-pass, frame-inside-clip, measured-within-proposal, siblings-disjoint, requirements-aggregate. check [RULE...] is a statement; tessera check NAME [SIZE] runs them over one app with no input. A violation names the rule and the node. siblings-disjoint looks through ForEach, if, and view groups to the container that places them, so layers a list builds under a ZStack are layers.
• Locators: layout= and id= join type=, text= (a substring of the leaf's own cells), and nth=. An empty value is a parse error.
• expect cost NAME<=N over bodies, bytes, measures, placements, and renders; the bytes come from a new package seam, TerminalSession.bytesDrawn (a running total, differenced per step).
• Trace v0: one JSONL record per step (event, drew, invalidation reasons, changed nodes and rows, cost). tessera run --trace DIR and TESSERA_TRACE write it; tessera replay FILE --app NAME runs it again and compares everything except bytes. Keys, pastes, resizes, and the terminal's focus changes replay; a mouse event or a terminal reply is refused by name.
• shot writes NAME.txt and NAME.grid.txt (the catalog's grid dialect). The grid never has more style rows than chars rows, and a screen with more than 62 distinct styles fails the shot as a harness error. NAME.txt now reads a wide grapheme as one character.
• Demos in DemoSupport, each with a desktop and a mobile scenario that shots both and asserts a frame (most also run check): text (with the archived Unicode samples), stacks (Spacer, Spacer(minLength:), a Spacer in a VStack, priority; absorbs the stack demo), frame-padding, zstack (layers from a ForEach and an if too), flex (two pages: the constraints, then a priority tier, a child with no .flex value, and cross-axis placement), foreach. The corpus is 16 scenarios.
• Library: ViewGraph.update() discarded a resize's viewport reasons, so a resize redraw reported none; fixed with a test and a line in the Slice 1 spec.
• Docs: the Developer tooling page, the scenario README, the tessera-devtools skill (0.3.0), mise-tasks/dev/{check,replay,run}, text.md (it says what the Text demo renders until Slice 8), the plan, and the CHANGELOG.

Critic review

Every finding is fixed in a7260a7, each behavioural one with a test: F1 siblings-disjoint through list nodes (crafted projections, live fixtures, the zstack demo and its scenarios; it fails without the fix), F2 the grid row rule, F3 62 styles fail closed, F4 run --pretty warnings, F5 empty arguments, F6 the demo gaps, F7 siblings-disjoint named in the Flex scenarios, F8 CHANGELOG, F9 the trace version first, F10 the help overview, F11 text= is a substring, F12 the ForEachDemo comment. The open question is settled by recording focus events replayably (Discoveries in the plan). The re-check's fix is a7d0b5c: a view list is recognised by its module-qualified type (TesseraCore.ForEach, Swift.Optional, …) and carried as the projection node's isList, so an app's own view named ForEach is not looked through (a live test fails under short-name matching), and the README says a container-less multi-view body is a true positive.

Verify

• mise run core:test: 959 tests green on Xcode 26.6 and on Xcode 27.1 (838 on the base).
• swift test --package-path Examples --only-use-versions-from-resolved-file: 39 tests green on both (25 on the base).
• mise run dev:run Examples/Scenarios: 16 scenarios, 16 passed; two runs give the same bytes, and --repeat 3 agrees.
• mise run quality:format, quality:lint, and docs:lint green, and the pre-commit hook's lint of Examples/Tests; no Package.resolved in the diff.
• CI on 6ea1bd5 (run 40087437615, after the critic fixes): all 14 jobs green, and every test leg printed 16 scenarios, 16 passed, 0 failed, Windows included. The head carries the run after the re-check's fix.

For the owner

Decisions and discoveries marked "(owner to weigh)" are in the plan. The ones most worth a look: a bare check fails on errors and only lists frame-inside-clip warnings (and run --pretty prints them); text= matches a substring of the topmost leaf; the grid golden's row rule and its 62-letter limit; live traces record focus events; and the library findings below.

Library findings, each recorded in the plan and left as today's behaviour unless noted:

• update() dropped a resize's viewport reasons (fixed here).
• Every step of an app with stacks reports layoutEnvironmentChanged and renderEnvironmentChanged: stacks rebuild their children's EnvironmentValues and reconciliation compares storage identity.
• The headless VT lacks DEC mode 2027, so ZWJ emoji and regional indicators advance different cells than Tessera's width table.
• A layoutPriority Spacer starves lower-priority rigid siblings.
• Flex's min(n) floor cannot compress while Text has min == ideal, fill weights invert when children are oversized, and Flex takes no alignment (a smaller child is always leading or top, though flex.md says "the selected" alignment).
• measured-within-proposal and frame-inside-clip fire by contract on a fixed frame wider than its parent, on Flex overflow, and on non-wrapping Text; the workspace demo at 40x16 trips them on a collapsed pane (Slice 8).
• A root @ViewBuilder body is always dirty, so every step draws (Slice 4 found it; the cost counters now show it).