[IanMayo]

Summary

Implements SRD Annex A Part A, spec 006 tasks T003 to T009. Stacked on #50: this branch contains #50's commit, so merge #50 first.

Browser-safe subpaths (T003, FR-009).

• New subpaths: @tma/features/browser, @tma/classical/browser, @tma/eval/browser and @tma/data/canonical. They re-export the existing modules with no forks.
• outputsToPrediction moves from models into features/decode.ts, and models re-exports it. It gains an optional first row, used by the viewer to show pre-zig rows. Goldens are unchanged.

@tma/demos core (T004).

• Catalogue types and scenario regeneration from a seed.
• Plan geometry, with a 1–2–5 grid between 100 m and 10 km.
• Time series and legs.
• Model inference glue that takes tfjs as a parameter.
• Classical rows: per-timestep, plus a whole-window batch track for time-late mode.
• Naive rows, per-row and scenario errors through @tma/eval, URL state, and NDJSON export.

Tests (T005).

• Browser bundle (FR-001): esbuild for the browser platform, run in a VM. Scenarios and classical rows are byte-identical to Node.
• Agreement (FR-007): tfjs CPU backend against tfjs-node on the 2c LSTM, 2c BiLSTM and i1 LSTM, to float32 tolerance.
• Golden view model and server render of all three modes from URLs.

React viewer (T006).

• Plan view with pan, zoom and a labelled grid. It shows the legs, turn and zig, headings, and observed bearing rays up to the cursor, with optional true bearings.
• Instantaneous mode adds the estimate trail, a velocity vector and a range ring. Time-late mode adds the model and classical tracks.
• Plots of range, course, speed, bearing and bearing rate on a shared cursor, with the gate points marked and the pre-zig period labelled unobservable.
• A readout at the cursor gives truth, model, classical and naive values with their errors. A summary gives the scenario's post-zig medians.
• Panels show the scenario metadata and the model pair. A note appears when the source is a fresh seed or lies outside the training ranges.
• Controls: dataset, index, previous, next, random (crypto), bookmarks (featured and saved locally), mode, model pair, overlays, play, and the time slider. PNG and NDJSON export.

Site (T007).

• The /viewer page loads the viewer browser-only and lazily, with a navbar link.
• Prebuild builds the catalogue from the four gate experiments' dataset headers, manifests and verdicts. The gate label is the panel decision, so iteration 1 shows "G1 pass".
• Prebuild copies the eight carried models (4.5 MB) into site/static/viewer/, which is git-ignored.
• site:build now runs tsc -b first, because the page imports @tma/demos from dist/.

Smoke and post (T008, T009).

• scripts/viewer-smoke.mjs is a Playwright smoke test of the served page. It passes here with no console errors and every pair runs.
• The 2c BiLSTM takes 1.08 s, against SC-001's 2 s target. The LSTM takes 0.46 s.
• Exports were checked in Chromium. The PNG renders. The NDJSON's scenario is byte-identical to the Node generator, and its predictions parse against the schema.
• The algorithm note site/docs/algorithms/viewer.md explains how the viewer works.
• The blog post links three views (SC-003): a typical case, the 2c LSTM's largest speed error at zig + 10 min (σ = 1°), and a held-out 26 km case.

Findings from the cases. These illustrate; they do not measure frequency.

• Jagged time-late track. The BiLSTM's positions form a jagged path, not a constant-velocity track.
• Worst speed case. The target opens nearly along the line of sight, so speed is weakly observable there. The classical solver also struggles: 13 of its 101 rows do not converge.
• Held-out range. On the 26 km held-out case, the model's range starts near 18 km, below the truth. Training ranges ended at 20 km.

Bug found and fixed. The site's Babel compiles [...nodeList] to [].concat(nodeList), which broke the PNG export. It now uses Array.from, and the pitfall is recorded in AGENTS.md.

T010, the owner's realism check, remains open.

Constitution check

• Data types defined only in schema/tma.yaml; pnpm schema:gen run (I): no schema change. The catalogue types describe viewer configuration, not pipeline records (plan, Constitution I).
• Deterministic: seeded PRNG only; golden files updated deliberately, with the reason stated (II): there is a new packages/demos/test/golden/views.ndjson, and other goldens are unchanged. Inference uses the CPU backend. Random scenario choice uses crypto.getRandomValues and affects no output.
• sim/features stay pure functions; one CLI per stage; Keras-equivalent layers only (III): they are reused through subpaths, and the decode function moved into features unchanged.
• Any reported result sits beside naive and classical numbers on the same test set (IV): every model figure in the viewer and the blog post is shown beside classical and naive.
• Respects the current gate: no training before G0, no scope beyond SRD §1 (V): nothing was trained or scored.

Site pages

• Glossary entry (site/docs/glossary.md) for any new term: no new term
• Data dictionary: regenerated automatically from the schema (nothing to edit)
• Algorithm note (site/docs/algorithms/viewer.md)
• Blog post (site/blog/2026-10-17-engagement-viewer.md)
• Not applicable: this PR adds none of the above

Checks

• pnpm lint && pnpm typecheck && pnpm test && pnpm schema:check && pnpm site:build: all pass, with 302 tests.
• node scripts/viewer-smoke.mjs against the built site: pass.

---

Generated by Claude Code

[github-actions]

Site preview: https://debrief.github.io/Gitma/pr/51/ (updates on every push; removed when the PR closes).