[Seungpyo1007]

What

Moves the TechEngine pin from ce09b14 to 0c5012b, and moves collect.yml's three uses: to the same commit.

The new pin brings in GetTechAPI/TechEngine#132. Ingest writes a value only when the source cell states it unambiguously:

• TDP. "45 W (35−54 W)" gives 45. A cell that is only a cTDP range ("12–25 W") now gives no value; it used to give the range's maximum.
• Several values in one cell (memory sizes, memory types, bus widths, clock or TDP ranges, "~" or "?") give no value.
• Dates. A day, a month or a quarter gives its first day. A half-year or a bare year gives no value.
• Laptop parts keep a laptop name (-mobile). A laptop-table row without one is skipped; it used to be written under the desktop card's name.
• Segment. Ryzen X parts are desktop, not HEDT. Chips under a Mobile heading are laptop.
• Ryzen PRO pairs. In a "( PRO ) 250" row, the cell "(March 27, 2025) February 18, 2025" now gives each model its own date.
• Repeats. Rows that share a slug but disagree on a value are all skipped. Smartphone variant records are matched by their base model.
• Two source pages that return 404 are dropped: List_of_AMD_Threadripper_processors and List_of_Xiaomi_smartphones.

Why now

ce09b14 stopped the duplicates, but a dry run on an empty data root still wrote wrong values. 127 of 469 CPU rows took a cTDP range's maximum as their TDP, every Ryzen X chip was HEDT, and GPU laptop rows took desktop names. On the same cached pages, the fields where 0c5012b disagrees with the curated record drop a lot:

Field	ce09b14	0c5012b
CPU tdp_w	131	32
CPU segment	124	54
CPU base_clock_ghz	16	5
GPU pcie_version	184	160
GPU memory_gb	63	30
GPU tdp_w	55	32
GPU memory_type	25	8

I checked each kind of change on the rows both versions write against its source cell. The desktop GTX 1070 now has the desktop card's clocks and TDP instead of the laptop's. Ryzen PRO pairs get their own dates and clocks, Phoenix and Hawk Point chips get the Zen 4 boost clock, and HX and Z2 chips get their Zen generation. Everything else that changed is a value now withheld or a socket written as LGA1151.

Checks

• ce09b14 is the first parent of 0c5012b, and the only change between them is that PR. It touches app/ingest, app/verify/wikipedia_gpu_backfill.py, their tests and docs/DATA_PIPELINE.md. ingest.yml, refresh.yml, verify-promote.yml, app/validate.py and the dump code are unchanged.

• The ingest and backfill tests pass on 0c5012b: 332 tests.

• Dry run of the merged collectors on a copy of develop's data, with every Wikipedia page fetched once:

  Category	Written	Already curated	Incomplete or uncertain
  gpu	1	1,494	310
  cpu	0	1,429	394
  smartphone	0	35	6

  • The GPU record is geforce-gtx-1660-ti-mobile, the GTX 1660 Ti (Laptop): 80 W, 1455/1590 MHz, 6 GB GDDR6 on 192 bits, released April 23, 2019. Each value matches the source row. Only the desktop card and the Max-Q version are curated.
  • Smartphone candidates fell because a row that names no single phone or carries no spec is no longer one. That covers captions, legend rows, "Pixel 2 / XL"-style pairs, and the Samsung list, which has no spec columns.

• pr-checks runs the full engine validator and a smoke dump, because the pin moves.

Closes #297