chore(server): promote the llama.cpp b10238 bump and dependency updates to production

[dvcdsys]

Promotes the current develop to main so server/v0.12.8 can be tagged.

What ships

llama.cpp b10238 (#217) — bumps every pin in the repo and fixes the bundle
breakage the new upstream layout exposed. fetch-llama.sh now ships the whole
*.dylib set instead of a hand-maintained list, matching the rule
Dockerfile/Dockerfile.cuda already encode as COPY /app/*.so*.

Two follow-up fixes found while regression-testing that change:

• 95e703d — the new @rpath guard could pass vacuously on exactly the bundle
  it exists to reject: otool exits 0 on a non-Mach-O file, and an absent
  otool fed the loop an empty list, either of which reads as "nothing
  missing". It now requires the tool, requires it to succeed, and rejects
  output that is not a dependency listing. Also corrects two docs claims the
  bump left stale (llama-checksums.txt asserted a CI gate that does not
  exist; README.md still documented the replaced fixed dylib list).
• 02ee183 — make bundle copied into a directory nothing ever cleaned, so
  the shipped bundle accumulated the dylibs of every version ever built there.
  A b8914-era tree rebuilt on b10238 produced 45 files where the fetch staged
  36, the extras being 9 stale b8914 dylibs (13 MB) sitting next to their
  replacements. Now 36 files / 52 MB, matching the fetch exactly.

Dependency updates — npm group (16 updates), Go group (3), GitHub Actions
(setup-go 7, setup-node 7, docker/login-action 4.5.1).

Verification

Full regression on all three build surfaces, each exercising the
llama-dependent path end to end (index a repo, then two semantic queries that
must resolve to different files — returning a hit is not enough, since a
broken embedding path still returns something):

Surface	Build	Run	Index	Search	Metadata
macOS native (b10238, Metal)	✅ 52 MB	✅ ready in 3.4s	✅	✅	✅
Docker CPU (arm64)	✅	✅	✅	✅	✅
Docker CUDA (amd64, RTX 3090)	✅ 1.48 GB	✅ healthy	✅	✅	✅

GPU attribution verified explicitly rather than inferred from container health:
the llama-server PID from docker top appears in
nvidia-smi --query-compute-apps with 504 MiB, ruling out a silent CPU
fallback.

Security gate — trivy diff of develop-cu128 against the released
v0.12.7-cu128, both scanned on the amd64 host:

base      v0.12.7-cu128 : CRITICAL=0  HIGH=4
candidate develop-cu128 : CRITICAL=0  HIGH=4
new: 0

All four HIGHs are the known pre-existing findings in the bundled ngrok
binary and are identical on both sides. This closes the gate PR #217 flagged as
outstanding.

🤖 Generated with Claude Code