Seafood stall inventory specialist - the live-stock estimation calculator: per-tank counts, size ranges, variance on the spot - recount while everyone is still standing at the tanks

Build me a single-file HTML live-stock count estimator that opens
in a phone browser, no network required.

Background: counting live stock = heads per tank per species x
average unit weight, then reconciliation against the books. Today
it runs on a pocket calculator, gets copied onto paper per tank,
and lands in a spreadsheet later with transcription errors - and
the variance rate only exists once everything is counted. By the
time a species shows a huge gap, everyone has gone home; no
recount is possible, and next week starts with a question nobody
can answer.

Input (per tank, form entry or CSV paste):
- Tank number / species / method (counted | eyeballed) / head
  count (eyeballed entries are ranges, e.g. 80-100) / average
  unit weight (defaults from that species' latest batch record,
  manually editable) / species purchase price
- Shellfish troughs: drained weight entered directly, tool
  converts to net via my [moisture coefficient] - I set it, the
  tool presets nothing

Computation:
- Per-tank estimate = count x unit weight (range counts produce
  range estimates); species totals, four-category totals, and
  the grand total accumulate automatically
- Book comparison: I paste the book expectation (species level);
  variance = (estimate - book) / book, at species and category
  level
- Species past my [alert line] highlight with a fixed message:
  "recount this species now" - the counter is still standing at
  the tank, and this moment is the cheapest a recount will ever
  be
- Eyeballed tanks display in their own block, never interleaved
  with counted-tank figures

Output:
- Count sheet (CSV export): tank / species / method / count /
  unit weight / estimate / book / variance rate
- Summary: one block per category + total variance rate, with
  the eyeballed block carrying its own line: "range estimates
  carry lower confidence than counted figures"

Hard requirements:
- Single file, no build step, no external CDN or fonts
- localStorage persistence - a locked phone screen mid-count
  loses nothing; JSON export and import
- No "smart correction": a big variance must never trigger an
  automatic estimate adjustment, only the recount prompt - to
  adjust or not is a human decision after the count; a tool that
  edits its own measurements is a forged count
- No preset species weight tables of any kind - regional and
  batch variation is so wide that "typical values" don't exist;
  only this store's own purchase records are real
- Eyeballed ranges keep both bounds all the way into the
  results; never display a "midpoint" as the answer - "roughly
  90" is more useful in variance analysis than "90," because
  the former is honest

Why it is written this way

The most expensive part of a live count is not the counting - it is discovering the variance too late. In the traditional flow the variance rate lands after everything is tallied; by the time a species shows 8%, the only person who could recount has gone home, and next week's investigation starts from a sheet nobody can re-verify, unable to say whether it was mis-estimated or truly missing. The tool's core function is a single move: pull the variance calculation forward to the moment when people still stand at the tanks, highlight the recount, and let ten minutes of on-the-spot verification save a full week of arguing.

"No automatic estimate correction" is a boundary of authority, not a technical problem. Models building tools love adding "smart calibration" - nudging estimates toward the book figure to shrink the variance. It looks thoughtful; it is target-value pollution of a measurement. A count's entire meaning is being an independent measurement from the books, and a count "corrected" by the books is no count at all. So the tool's mandate is strictly "compute, compare, prompt" - the authority to touch the numbers stays human.

Keeping range bounds unblended treats honest uncertainty as part of the data. Eighty-to-one-hundred estimated as a range stays a range; displaying "90" is pretending certainty. In variance analysis, wide-range figures deserve lower weight than narrow ones - information that exists only as long as the range does. A count's credibility never comes from pretty numbers; it comes from every number declaring how much of itself to trust.

No preset weight tables follows the same discipline as the other tools: "three and a half jin" for grass carp is a common value in some regions and seasons, and applied to another stall it becomes systematic bias. Size data recognizes only this store's purchase records; the tool leaves its cells empty and waits for data rather than imagining it for the user.

Variables