Parse transferable days from FK instead of deriving them

The scraper never read "överlåtbara dagar" from Försäkringskassan. It only
visited ansok-och-planera-foraldrapenning#/valjdagar and derived the
reserved/transferable split in build_snapshot, assuming transferable days
are spent before reserved ones. On real data that gave 49 for the user and
0 for the partner where FK states 129 and 71.

The assumption cannot be repaired: reserved-first, transferable-first,
proportional and cap-based models were all checked against real figures and
none reproduces FK's accounting (reserved-first fits the partner but is 10
days off for the user). FK publishes the real numbers per parent per tier on
the min-foraldrapenning page, which the scraper had never visited.

- Scrape the "Dagar ni kan föra över" panel after the valjdagar extraction,
  summing sjukpenningnivå + lägstanivå per parent. The totals are not
  literals on the page. Parse only direct-child <p> elements: the first
  group embeds a tooltip with decoy "45 dagar"/"90 dagar" text.
- build_snapshot accepts the reported value and derives reserved days from
  it; the old arithmetic remains only as a fallback. An implausible FK value
  degrades to that fallback rather than failing the fetch.
- Record provenance in transferable_days_reported (migration 006) and badge
  derived values as estimates in the dashboard breakdown. Drop the footnote's
  claim that at most 150 days are transferable — real data contradicts it as
  a per-family figure.
- Select the child's section by name when FK lists several; never guess.

Verified against a live BankID scrape: scraped 240/101/139 with 129
transferable, deduced partner 240/169/71 with 71, both flagged as reported.

Incidental fixes found along the way:

- The fetch router never forwarded the partner's transferable figure to
  persist_partner_deduction, so the partner's stored snapshot would have kept
  the flawed split even once the parse landed.
- 'deduced' snapshots were labelled "Manually entered" by a two-way ternary
  over a three-value union. Replaced with a compiler-checked exhaustive map.
- Alembic resolved its database from alembic.ini alone, ignoring
  APP_SQLITE_PATH, so migrations silently targeted the developer's real
  database even when pointed elsewhere. env.py now resolves it, preferring
  -x sqlalchemy.url, then a real APP_SQLITE_PATH, then the ini value.
- backend/captures/ was not ignored: the root pattern was anchored as
  /captures/ despite a comment saying scraper dumps must never be tracked.
- CLAUDE.md quoted test counts and coverage percentages that had gone stale,
  which made a correct verification report look suspect. Gate thresholds kept.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DXPBfr1v7F71t6QcbHCaaT

CLAUDE.md

@@ -114,7 +114,7 @@ uv run ruff check .              # lint
 uv run ruff format --check .     # format check
 uv run pytest                    # tests + coverage (config in pyproject.toml)
 ```
-Gate: `--cov-fail-under=80` (total now ~84%), climbing toward **90**. Never lower it; raise only as coverage grows.
+Gate: `--cov-fail-under=80`, climbing toward **90**. Never lower it; raise only as coverage grows.
 
 **Frontend** (from `frontend/`):
 ```bash
@@ -123,7 +123,7 @@ npm run lint                     # ESLint, zero warnings
 npm run coverage                 # Vitest with coverage
 ```
 Gates (baseline measured after including `src/pages/**`): lines 96 | functions 91 | branches 92 | statements 96.
-A regression ratchet, not a target. Current: 97.26% lines/statements, 93.05% functions, 93.51% branches — above thresholds. Never lower gates to pass.
+A regression ratchet, not a target. Verify current coverage is above these thresholds; never lower gates to pass.
 
 ### WHO RUNS VERIFICATION (cost rule)
 
@@ -136,7 +136,7 @@ These checks are mechanical: fixed commands, judgement limited to reading pass/f
 Applies to full-suite runs: `uv run pytest`, `uv run ruff check .`, `uv run ruff format --check .`, `npm run typecheck`, `npm run lint`, `npm run coverage`, `npx vitest run`, `uv run alembic upgrade head`.
 Does **not** apply to quick mid-task diagnostics (reading one file, a `git grep`, one focused test) — delegating those costs more than it saves. The rule targets long whole-suite passes.
 
-**Brief the Haiku verifier with:** exact commands, directory to run from, expected baseline (e.g. "currently 361 passed / 0 failed — must not regress"), and an instruction to paste **raw output verbatim**, failures included, not a summary. "All green" without output is useless — several agents have reported gates passing that were failing.
+**Brief the Haiku verifier with:** exact commands, directory to run from, and an instruction to report exact pass/fail counts and coverage percentages from their run, then compare them against the current branch state to verify no regression; require **raw output verbatim**, failures included, not a summary. "All green" without output is useless — several agents have reported gates passing that were failing.
 
 ### MODEL TIERS