new(progress-report): replace build-plan.html with a progress report

[MendixMau]

What

Replaces architecture/build-plan.html with a generated progress report (architecture/progress-report.html, self-contained, stdlib Python behind build-plan-status.sh). It keeps everything the old page showed (phases, built and proven views) and adds stages P to 6 with artifact links, a progress bar, a sticky stage nav, and, when a Claude Code session log is available, time per stage, sub-agent use, context size, an estimated token and cost figure, a run health strip, a speed panel and a right-hand column of actionable tips.

• Session logs are optional. Without one the project sections render and time and token data are marked unavailable. Only numbers, stage names, script names, artifact paths and short sub-agent descriptions are read; no prompt or message text is copied.
• Tips are grouped by theme, collapsed by default, with calm severity. Costs inherent to the toolkit or mxcli (builds, required gates, waiting) are counted but never shown.
• Context window is read per model from a small table and raised if the log shows more; never a percentage over 100.
• Setting Progress report: off | ask | auto in PROJECT.md (default ask). Regenerated on every exec.sh slice; the 30 s self-reload applies only while a build is active.
• An existing build-plan.html is still accepted by gate-check and status.sh and never deleted. --json and build-plan.json are unchanged.

Impact

Row	Answer
Process	Surface artifact for stage 4 is now progress-report.html; build-plan.html still accepted. New optional PROJECT.md setting.
Dependencies	None new. python3 stdlib only; page skipped when python3 is absent.
Context cost	render-routing.sh --check: base 75204 words, head 75204 words (budget 80000). Skill words total about 308277 vs 308384 on master.
Behaviour	RUN on a real in-progress project (Stage 5, 153 plan rows) with its real session log: stage switch from the first metrics-ledger row, Stage 4 and Stage 5 windows match the chart band and the sub-agent tables, last build step matches the newest ledger row. Fixture tests cover the rest (F10 to F14).
Runtime	Generation is a single Python pass over the session log; no network.

Tests

• tests/wave2/test-progress-report.sh (new): 39 checks on a synthetic project and synthetic JSONL log.
• tests/wave2/test-build-plan-json.sh: 66 pass. tests/wave2/test-surface-column.sh: 13 pass.
• bin/check-portability.sh on changed files and render-routing.sh --check: clean.

Not verified

Visual layout (sticky side column, phone order, severity dots) was not viewed in a browser. Full test suite and CI not run locally.

Round 5 measured (head 0a908a3)

• Generation time on the real project above with an explicit session log: 1 s. With discovery against a 300-file ~/.claude/projects: about 1 s cold, under 1 s warm (discovery cached in .claude/progress-report-sessions.txt; F13 asserts a cached run does not reopen session files).
• Autoreload keeps the reader's place: scroll position and open sections are saved in sessionStorage and restored, history.scrollRestoration='manual', no reload within 20 s of interaction, and a Live / Pause control (F14).
• Tests: test-progress-report-field-fixes 90, test-progress-report 101, test-build-plan-json 70, test-build-plan-status-args 6, all passing.

🤖 Generated with Claude Code