Task 3 complete. Repaired working copy delivered in `output/site/`; original inputs untouched.

**Fixes** (cited to `inputs/brief.md` and supplied `site/` files):
- `app.js`: endpoint `/api/lead` → `/api/leads` (per `server.py` `do_POST`); discount formula corrected to `subtotal × (1 − rate) × 1.12`; added `parseSeats` (whole numbers 1–8 only) so blank/0/negative/fractional/9+ show "Enter 1–8 seats" instead of a price; client-side name/email/seats validation mirroring server rules, specific error messages, server-400 message surfacing, in-flight button disable, reset on success.
- `style.css`: fixed `width:1120px` → `max-width` + margins; single-column layout under 820px. Dark-blue/orange identity, title and all copy preserved.
- `server.py`: unchanged, verified byte-identical to input.

**Checks performed** (logs in `output/evidence/`):
- 19 browser sessions at 1280px and 390px — all `overflow:false`, no page errors.
- All four brief calculator examples matched in-page (268.80 / 537.60 / 725.76 / 1935.36); five invalid seat inputs rejected.
- 2 valid submissions (both widths) → visible confirmation and exactly 2 records in `submissions.jsonl`; 4 invalid submissions → errors, zero records created.
- Python: pricing/parse logic verified vs brief; server-copy identity check.

**Limitations:** Python sandbox blocks loopback sockets, so `server.py` itself wasn't exercised over HTTP — server behavior observed via the browser tool's trusted `/api/leads` implementation. Browser evidence is textual, not screenshots; mid-range widths and screen-reader flow untested.

**Outputs:** `output/site/{index.html,app.js,style.css,server.py,submissions.jsonl(empty)}`, `output/handoff.md`, `output/evidence/{browser-checks.md,python-checks.txt,submissions-after-tests.jsonl}`, `checks/test_server_and_math.py`.

**Estimated human finish time:** ~10–20 minutes (estimate, not measured) — live `server.py` run, visual eyeball, optional mid-width/accessibility pass.