The repaired site is in `output/site/`: signups now store through the unmodified supplied server, the calculator matches all four brief examples, and there is no horizontal overflow at 390 px or 1280 px. I did not run `python3 server.py` over a real socket, because the sandbox blocks socket binds from Python. Storage was confirmed through the browser tool's test server and by running the supplied handler in-process.

## What was wrong and what changed
- **Endpoint:** the form posted to `/api/lead`, but the server only accepts `/api/leads`, so every signup returned 404. It now posts to `/api/leads`.
- **Calculator:** the formula multiplied by the discount rate, so one seat showed CAD 0.00. It now takes 10% off for 3+ seats and adds 12% tax afterwards.
- **Seat validation:** blank, 0, negative, fractional and 9+ now show `CAD —` and "Enter a whole number of seats from 1 to 8."
- **Mobile layout:** the fixed 1120 px shell overflowed at 390 px. The shell is now fluid, with a single column at 760 px and below.
- **Form errors:** invalid submissions now show per-field messages and "Signup not sent…", and nothing is sent to the server.
- **Preserved:** title, all page text and the dark-blue/orange palette are unchanged. `server.py` is byte-identical to the input.

## Checks performed
- **Layout:** the original site reported overflow at 390 px; the repaired site reported none at 390 or 1280 in every run, with no JavaScript errors.
- **Calculator in the browser:** 1 seat = CAD 268.80, 2 = 537.60, 3 = 725.76, 8 = 1935.36. Blank, 0, -1, 1.5 and 9 were all rejected.
- **Valid submissions:** three, at both widths, each showed "Thanks! Your workshop request is saved." with a seat and total summary.
- **Invalid submissions:** four (empty form, bad email, 9 seats, blank name with 1.5 seats) each showed errors.
- **Stored records:** the test server's `submissions.jsonl` held exactly the three valid records and none of the invalid ones. I moved it to `output/evidence/` so the site ships without test data.
- **Supplied handler, in-process:** valid payloads returned 201 and stored one record each; `/api/lead` returned 404; five invalid payloads returned 400 with no record.

## Remaining limitations
- **No visual inspection:** the browser tool returns text, not screenshots, so colours, spacing and tap-target size were not viewed.
- **Not exercised in the browser:** the server-side 400 message, the network-failure message, non-numeric seat text, seats 4–7, and keyboard or screen-reader use.
- **Privacy risk in the supplied server, left unchanged:** it serves its whole directory, so from reading the code `GET /submissions.jsonl` would expose stored names and emails. It binds to 127.0.0.1 only. This needs a decision before any non-local use.
- **Duplicates:** the server stores repeat submissions as separate records.

Estimated human time to finish: about 15–20 minutes to run the server locally, view the page at both widths, and decide on the privacy point.

## Files in `output/`
- `handoff.md`
- `site/index.html`
- `site/style.css`
- `site/app.js`
- `site/server.py`
- `evidence/browser_test_submissions.jsonl`
- `evidence/server_check.py`