fix(data): merge 28 confirmed PDA near-duplicate records

[Seungpyo1007]

What

Near-duplicate scan of the pda category (140 records — small enough to review
every match by hand). Merged 28 confirmed duplicate records into 21 keepers
(140 → 112 files), keeping the better-sourced side field by field and unioning
source_urls. Anything ambiguous was left unmerged and is documented below.

Both gates pass on the merged tree:

• python -m app.validate → passed
• integrity_check.py <data> --strict → 0 hard anomalies (structural section empty)

Per-pattern breakdown

The same three patterns the tablet/watch (#324) and smartphone (#329/#330) passes
turned up all appear here.

Pattern 1 — identical scrape imported twice under different row-id slugs

Two cigarplug/scrape-gsma rows for the same device, byte-identical except
source_row. Kept the lower row-id, deleted the other (no field change needed):

• motorola/2002/accompli-388 — kept …-scrapegsma-344, deleted …-1227
• mitac/2005/mio-a701 — kept …-scrapegsma-278, deleted …-1161

Pattern 2 — same device under a brand-prefixed slug and a bare slug (sibling dirs)

Same convention as the #330 GSMArena/phonedb merge: kept the bare GSMArena/Kaggle
record (real specs + raw_gsmarena, sometimes Wikipedia), upgraded name to the
brand-inclusive form, unioned source_urls to retain the phonedb reference, kept
the keeper's soc/release_date, left verified as-is, deleted the phonedb twin:

Keeper (kept)	Deleted twin	New name
treo-680	palm-treo-680-palm-lowrider	Palm Treo 680
treo-750v	palm-treo-750v-htc-cheetah	Palm Treo 750v
treo-750	palm-treo-750-htc-cheetah	Palm Treo 750
treo-500v	palm-treo-500v-palm-otto	Palm Treo 500v

Treo 750 vs 750v and Treo 500 vs 500v are distinct SKUs — only the
matching-model twins were merged. palm-treo-500-palm-otto (2008 Treo 500) is a
different device and was left untouched.

Pattern 3 — cross-source duplicate (clean short-slug record + scrape-gsma row, same base dir)

Kept the clean record (correct soc FK, canonical slug), deleted the scrape twin,
and repaired the keeper's ram_gb from the twin's cited memory_internal — the
clean side carried the fabricated ram_gb = 1 placeholder while the scrape side had
the real figure. github.com/cigarplug/scrape-gsma was unioned so the repaired RAM
is sourced. No value invented: in every pair the deleted twin already held the real
value.

• 16 HP iPAQ pairs — h6310/h6315/h6320/h6325/hw6510/hw6515/hw6910/hw6915/rw6815/
  rw6818/rw6828/514 → 0.0625 GB (64 MB); 610c/910c/data-messenger/voice-messenger
  → 0.125 GB (128 MB).
• i-mate — pda2 (two identical scrape twins -1004/-121) → 0.125 GB;
  pda2k → 0.125 GB; pdal → 0.062 GB.
• palm treo-pro (3-way: bare Wikipedia-cited + …-scrapegsma-7066 + phonedb
  palm-treo-pro) — kept the bare record, ram_gb 1 → 0.125 GB (128 MB), name →
  "Palm Treo Pro", unioned all four source_urls.

Left unmerged (ambiguous — documented, not guessed)

• palm-pre-scrapegsma-7065 (GSM webOS Pre, 2009-10) and palm-pre-plus-scrapegsma-7063
  (GSM Pre Plus, 2010-03) overlap a region-split family — palm-pre[-plus]-palm-castle-c
  (CDMA) and palm-pre[-plus]-gsm-{eu,na}-palm-castle-g — but do not map 1:1, so
  no merge was made.
• The EU/NA regional variant pairs (palm-pixi-gsm, palm-pixi-plus-gsm,
  palm-pre-2-gsm, palm-pre[-plus]-gsm-…-castle-g) are legitimate separate SKUs
  (different release dates + regions) and were left alone.

Systemic finding (reported, not mass-edited)

The same fabricated-defaults signature the earlier passes flagged is present here:
nearly every clean pda record carries ram_gb = 1 (real values are 16–256 MB),
and many Palm records carry os = "Feature phone" for what are Palm OS / webOS
devices. I only repaired ram_gb inside the confirmed merges (where the deleted
twin supplied a cited real value); the remaining placeholder records were left
untouched rather than mass-edited, since fixing them needs a per-record source and
is out of scope for a dedup pass (per the "never invent values" rule).

Ownership / dump

data/pda/ only, plus the mandatory dump refresh as the last commit (re-dumped the
pdas collection from a fresh scratch DB with TECHAPI_DATA_DIR on this worktree;
pdas 140 → 112, no other collection touched).

Closes #296 — this clears the pda slice of the "near-duplicates in other categories
not yet scanned" item, but that item stays open (smartphone still has 322
documented-unresolved pairs per #330, and other categories are not yet scanned), so
this is Refs, not Closes.