TSO 500 case report: manual pass over the printed document (HTML/PDF)

🤖 Written by Claude.

The printed case report (HTML preview + PDF) has had seven content changes land against #444 with unit
coverage over synthetic contexts only. Everything below is visible by building **one** case report and
reading it — the Word file off the same build is #1874, the JSON is
SACGF/variantgrid_sapath#448.

All of it comes off one `ReportContext` (`classification/report/case_report_context.py`), so a content
bug here shows in the JSON too — check the document, and let the JSON issue check the contract.

Test against the **SA Path TSO 500 template** — it is the only user of multi-variant reporting, so there is
no need to exercise the generic template.

Build one, then: `/classification/case_report/` for the preview, `.../download/pdf` for the PDF.

## Results Summary

- [ ] A gene-level `` / copy number loss prints in its own **Copy Number Losses** table, not beside
the gains in Copy Number Changes. Both tables print even when empty ("None detected") — all five
kinds are always in `kind_groups`, so a skipped kind never reads as "not looked for".
- [ ] Within Copy Number Losses, **fewest copies first** (CN 0 above CN 1) — a loss's biggest finding is
the lowest count. Within Copy Number Changes, most copies first, as before.
- [ ] A call with no copy number at all sorts behind the ones that have one, in both tables.
- [ ] VAF prints as a whole percent rounded **half up**: a 12.5% VAF is **13%**, not 12. (Python's own
rounding is half-to-even and would give 12; the value is rounded once, in `vaf_percent_whole`.)

## Splice note

The sentence above Variant Interpretation, generated rather than typed.

- [ ] One reported small variant with protein designation `p.(?)` → "Note that the **X** variant with
protein designation p.(?) is an intronic variant that is predicted to disrupt splicing, see below."
- [ ] Two or more → "...the **X, Y and Z** variants ... **are** intronic variants ..." (comma list, "and"
before the last).
- [ ] **TERT is excluded**: a case whose only `p.(?)` variant is TERT produces *no* note — its promoter
variants are non-coding for a different reason.
- [ ] A not-reported `p.(?)` variant does not put its gene in the note.
- [ ] Cross-check: the genes named in the note match the genes the curator named in the JSON's
`Mutations Comment`. These are **not** fed from each other — the note is generated, the comment is a
typed case value — so if they disagree in practice we need to decide which wins.

## Tier sections

Behaviour change: a tier is now dropped on having nothing **reported**, where it used to be dropped on
having nothing in it at all.

- [ ] Tier I, Tier II and Tier III print even when empty (`ALWAYS_PRINTED_TIERS`).
- [ ] **A case with a Tier IV variant shows no Tier IV heading.** A Tier IV variant is a not-reported
record, so the section would otherwise be a heading over "no reportable variants detected" on every
case that has one. This is the main thing to confirm with the lab.
- [ ] Same for "Tier I or II" and "Not tiered" — heading only when something in them is reported.
- [ ] `unreported_count` on a printed tier still counts the records held back.

## Gene-level gene symbol

- [ ] A gene-level CNV row prints the **approved** symbol — a caller's `MYCL1` reports as `MYCL`
(see the `SEGID=MYCL1` check on SACGF/variantgrid_sapath#446).
- [ ] A fusion prints its **anchor** (first) gene as the symbol, matching the representative annotation.
- [ ] Small variants are unaffected — their symbol still comes off the transcript.

## DRAFT watermark (PDF)

- [ ] A preview and any non-finalised report carries DRAFT.
- [ ] **xhtml2pdf implements neither `transform` nor `position`**, so the PDF's DRAFT sits in the document
flow rather than rotated behind the text. Confirm the lab accepts that, since it's a design call we
can't fix in xhtml2pdf.
- [ ] After finalise, the stored `pdf_file` has **no** watermark — `finalise_case_report` now re-renders
the documents. Download the PDF again after finalising rather than trusting the preview.
- [ ] "Rebuild documents" on a still-draft report **keeps** the watermark.
- [ ] Finalise re-renders over the stored snapshot, so no number the report already quoted moves.