#3445 [Other] update the Coder model pin to the 2026-10-19 release

Description

## Problem and result

ISTA replaced the Coder's **first shard** (the expert-pruned model) on 2026-10-19 and updated the expert allocation to match it. `setup.py` still pins the 2026-09-29 revision, so a fresh setup installs the older build: SWE-bench Verified **75.60** instead of **79.20**, LiveCodeBench v6 **86.28** instead of **87.43**, and the looping the release reduces is still there.

With this change the pin names the new revision and the new shard-1 SHA-256, so setup downloads the improved model. The n-gram table (shard 2) and the picture reader are byte-identical upstream, so their pins stay as they are.

## Related discussion

none

## Approach

Two lines, both in `setup.py`:

- `HF_REVISIONS["ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF"]`: `5348543e…` -> `bba88856f8c3c0c9e40e45f82d05e228b814ac43`
- `CODER_FILES["…-00001-of-00002.gguf"]`: the SHA-256 `e11083ba…` -> `bfa3644db4adfff072fc946c20994f2909d2cf89879d83f80a4c811d4016c211`

Nothing else changes: setup still checks the size and the SHA-256 of every file, and the model files themselves are not touched.

## Validation

**Passed:**

- `sha256sum -c` of the downloaded shard 1 against the new pin: `OK` (29,608,464,996 bytes).
- `tools/iq_pack.py --gguf <new shard 1> --out <pack> --compat-bf16 --experts-bin` built a pack from it: `experts.bin` 25,146,163,200 bytes, the same layout as the previous pack (`native_experts.txt` and `index.txt` are byte-identical, only the expert bytes differ).
- The engine started from that pack and served: 5/5 code prompts answered, a 20,016-token prompt read at 1,477 tok/s, 35-52 tok/s decode.
- The pack tool needs both shards as siblings; shard 2 was hard-linked from the existing copy (same bytes, `sha256 316b46f3…` unchanged), so nothing was downloaded twice.

**Failed:**

- none observed

**Not run:**

- Windows setup, the app and the Terminal (the change is two lines in `setup.py`; the same pin is read by every install path).
- HIP, SYCL and Metal builds (no engine code is touched).

## Limits

- The expert profile Strata ships for the Coder (`data/expert-profile-coder.bin`) was mapped onto the kept experts through the release's **old** `rco-allocation.txt`, so it no longer describes this build: only 2 of its 12,288 ranked pairs still point at the same expert. It costs little at runtime, because the engine orders the expert cache by its own learned heat and the shipped profile is only the tie-breaker (`src/core/expert_cache.cpp:160-164`), but a profile re-mapped for the new allocation would be better. I can post one if it is useful.