Recover orphan text after band normalization (TSR stage 2), v1.6.2 (#63)

PR #62 (1.6.1) closed the cell-bleed regression by clamping SLANet's
loose cell bboxes into non-overlapping row/column bands and tightening
text membership to center-containment OR >=60% overlap. That worked,
but exposed the opposite failure: legitimate native PDF text whose
center fell just outside the *clamped* cell bbox now had nowhere to go.

Two distinct failure modes observed against the FNBO branch-list PDF
after 1.6.1 deployed:

  Symptom A — header text positioned at the LEFT of a column whose band
  was derived from data-cell centers farther right. Header "Address"
  PDF text at x=331..375 fell outside the clamped col band starting
  at x=410. Strict membership rejected it (0% x-overlap, center
  outside).

  Symptom B — local SLANet row drift in col 0 over a 5-row stretch.
  Cell bboxes sat just above the actual branch-name text items
  (x-overlap 100% but y-overlap ~30-43%, below the 60% threshold).

Both share one root: post-normalization bboxes are too tight, and the
strict rule has no escape valve for legitimate edge text.

Fix: add a stage-2 orphan-recovery pass after the strict fill. Items
that NO cell claimed in stage 1 get re-assigned to their nearest
*empty* cell, distance-capped by `(median_col_width, median_row_height)`
so a far-orphan figure title can't get pulled into a faraway empty
cell. Stage 2 only fills empties — never overwrites stage 1 — so the
cell-bleed case PR #62 closed cannot regress.

Three new lib tests cover the bug shapes:
  - stage2_recovers_left_aligned_header_text_outside_data_band
    (Symptom A: header text left-of-band, data cells already filled,
     stage 2 fills only the header)
  - stage2_recovers_y_shifted_col0_in_consecutive_rows
    (Symptom B: 3 col-0 cells shifted vs text, all 3 recovered)
  - stage2_does_not_overwrite_filled_cells_or_admit_far_orphans
    (cap rejects far figure titles; pre-filled cells untouched)

Plus tests for the cap-derivation helper:
  - tsr_assignment_caps_uses_median_geometry
  - tsr_assignment_caps_floor_protects_degenerate_input

The existing dense-overlapping-rows regression test (added in #62) still
passes, confirming no regression on the cell-bleed case.

Test results: 414 lib + 115 integration + 2 doctests pass. clippy + fmt
clean.

Bump @firecrawl/pdf-inspector to 1.6.2 (patch — fixes a regression
introduced by 1.6.1; no API changes).

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

napi/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@firecrawl/pdf-inspector",
-  "version": "1.6.1",
+  "version": "1.6.2",
   "description": "Fast PDF classification and text extraction. Detect text-based vs scanned PDFs, extract text by region with quality checks. Native Rust performance via napi-rs.",
   "main": "index.js",
   "types": "index.d.ts",