test(ArcGate): fail CI when a feature is built, wired, and unreachable

[heydryft]

Jish's standing order: "its built and its switched off should never happen again." This is the structural half — the guard that makes it fail CI instead of being noticed months later.

Why nothing catches this today

mistralrs-core is deliberately outside the -D warnings clippy lane (.github/workflows/ci.yml:131 — the workspace-wide form floods with pre-existing upstream findings). So the one signal the compiler does emit for this class, dead_code, is discarded. A feature can land wired-to-nothing and the only trace is a warning nobody gates on.

What changed

One new file, mistralrs-core/tests/capability_reachability.rs. It runs in the existing cargo test -p mistralrs-core lane (ci.yml:88) on ubuntu/windows/macOS — no new workflow, no GPU, no model.

check	what it asserts
every_registered_capability_has_a_production_call_path	each REGISTRY entry has a reference outside tests. Tracked entries are known-dark with a reason and go red if they become reachable — the list cannot rot in either direction
no_new_unreachable_symbols_in_arc_owned_code	the ratchet: any Arc-owned definition with no production reference fails unless it is in DEAD_SYMBOL_BASELINE
9 further tests	the scanner's own self-tests (below)

Check::ApiPromise encodes the logit_bias class the coordinator asked for: a parameter accepted at the API boundary that must be read by the runtime meant to honour it. Accepting a parameter and silently dropping it is worse than rejecting it — the caller believes it took effect.

The guard was proven able to fail, three ways

1. Live acceptance against PR #125, not a synthetic case. prefill_chunk_is_intermediate (pipeline/mod.rs:421) is introduced by feat/chunked-prefill-cursor with zero callers — not even tests. Copying this file onto that branch:

1 NEW switched-off symbol(s) in Arc-owned code:
  mistralrs-core/src/pipeline/mod.rs:421  `prefill_chunk_is_intermediate` — defined, never referenced by production code

Exactly one symbol, correct file:line, no noise from the rest of that branch. Not fixed here — that is #125's author's call.

2. A hidden live call site. batch_can_be_ragged has one production caller (kv_cache/mod.rs:1520) and four test callers. Inlining the production one left the four test callers in place; both checks went red independently. This proves the production/test discrimination, not merely symbol presence. Reverted.

3. The guard caught itself passing vacuously. On its first run the logit_bias entry was green — on a tree where the wire is dead. logits_bias occurs in sampler.rs exactly twice, as the declaration (:85) and its None default (:124), and both were counted as reads. Fixed by reads_field, which excludes any occurrence immediately followed by : (declaration or struct-literal write). The entry then correctly reported:

DEAD WIRE: `sampling: logit_bias reaches the sampler` (ArcInfer / ArcSample)
    field `logits_bias` IS accepted at the API boundary (chat_completion.rs:630, completions.rs:239)
    but is never read in ["mistralrs-core/src/sampler.rs"].

It is registered Tracked naming PR #151, which implements it. When #151 lands this test goes red asking to be promoted to Live — the bidirectional design doing its job rather than the fix being forgotten.

Two scanner bugs were also found by running it, both of which under-reported test regions and would have mislabelled test code as production: a fixed lookahead window missed #[cfg(test)] separated from its mod by a comment block, and whole-file test modules (#[cfg(test)] mod parallel_bake_tests;) were not detected at all.

Baseline — five pre-existing dark symbols, now tracked rather than forgotten

Cross-checked against cargo check -p mistralrs-core --message-format=json at cca5e5c6e; rustc's own dead_code analysis independently reports three of them.

file:line	symbol	note
arc-cuda-graph/src/decode_forward.rs:115	arc_launch_gemv_bf16_silu_mul_down	a fused SiLU·mul·down GEMV compiled into libarccudagraph.a on every CUDA build and launched by nothing; declared twice, called from neither
mistralrs-core/src/kv_cache/xs_rolling.rs:516	assign_row_lens	✅ rustc agrees; only caller is inside a #[cfg(test)] region
mistralrs-core/src/pipeline/mod.rs:588	cuda_graph_runner_mut	a Pipeline method implemented by all eight pipelines and called by nobody
mistralrs-core/src/pipeline/mtp_pipeline.rs:1761	extend_draft_kv	✅ rustc agrees
mistralrs-core/src/pipeline/mtp_pipeline.rs:2527	run_target_forward	✅ rustc agrees; only the batch variant is live

Also registered Tracked: AutonomousDecodeConfig (arc-cuda-graph/src/autonomous.rs:70-82) carries temperature, top_p, both penalties and greedy — and neither top_k nor min_p. Latent rather than live only because that path is itself unreachable today.

Scope — deliberately not included

The default flips are held, and the reasons are evidence, not caution:

• Ragged decode for V4 stays opt-in. PR feat(ArcSched): --v4-ragged-decode — V4's ragged decode path had no surface but an env var #153 already gives it a real CLI flag, and names a live correctness defect on master: a ragged cohort masked from one shared query position lets shorter rows attend compressed blocks they have not reached — silently, no error. That is fixed by fix/ragged-gate-coherence, which must land first. Flipping the default now would ship a silent wrong answer. D14 also bans a CPU-only correctness verdict and this path has never run on a GPU.
• ARC_PREFILL_MAX_SEQS / ARC_PREFILL_FLOOR_STEPS deserve CLI flags but mistralrs-cli/src/args/mod.rs is being rewritten by feat(cli): cut the menu to 12 flags — opinionated surface, --help-all for the rest (D18) #110 and feat(ArcSched): --v4-ragged-decode — V4's ragged decode path had no surface but an env var #153 concurrently; adding flags here would conflict textually for no benefit.

Known limitation, stated rather than hidden

Reachability here is "a production reference exists", not "that reference is itself reachable" — a cluster of dead functions calling only each other satisfies check 1. It also cannot see env-gated-but-default-off features, which is most of the inventory; those need the CLI-flag work above.

How verified

No GPU; the GPU box was not touched.

• cargo test -p mistralrs-core --test capability_reachability → 11 passed, 0 failed at cca5e5c6e.
• Fail-before/pass-after observed for all three failure modes above, each output quoted verbatim.
• rustfmt --edition 2021 on the single new file only (never --all, per fork policy).

[github-actions]

Code Metrics Report

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Language              Files        Lines         Code     Comments       Blanks
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 C Header                  5          305          210           52           43
 CSS                       2         1181         1036           34          111
 CUDA                     72        24645        17702         4215         2728
 Dockerfile                1           39           22            8            9
 JavaScript               16         3546         2676          482          388
 Jinja2                    7          694          656            5           33
 JSON                     74         4600         4597            0            3
 Makefile                  1            6            5            0            1
 Metal Shading Lan|       33        12224         9431         1142         1651
 PowerShell                1          300          227           30           43
 Python                  144        14955        12317          810         1828
 Shell                    26         6312         4307         1564          441
 Plain Text                4         3801            0         2479         1322
 TOML                     33         1488         1293           45          150
 YAML                      3           25           23            2            0
─────────────────────────────────────────────────────────────────────────────────
 HTML                      4         2687         2604           43           40
 |- CSS                    2          543          479           37           27
 |- JavaScript             1         1233         1215           12            6
 (Total)                             4463         4298           92           73
─────────────────────────────────────────────────────────────────────────────────
 Jupyter Notebooks         4          122           83           23           16
 |- Markdown               1           60           30           22            8
 |- Python                 1          122          113            1            8
 (Total)                              304          226           46           32
─────────────────────────────────────────────────────────────────────────────────
 Markdown                197        39646            0        30479         9167
 |- BASH                  72         1652         1200          331          121
 |- C                      3           17           17            0            0
 |- CUDA                   2           84           56           16           12
 |- JSON                  18          708          708            0            0
 |- PowerShell             1            1            1            0            0
 |- Python                23         1008          787          113          108
 |- Rust                  65         2048         1713           77          258
 |- TOML                   6          207          164            0           43
 |- YAML                   4           38           33            5            0
 (Total)                            45409         4679        31021         9709
─────────────────────────────────────────────────────────────────────────────────
 Rust                    668       318690       275524        15018        28148
 |- Markdown             484        26172          471        22580         3121
 (Total)                           344862       275995        37598        31269
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Total                  1295       469159       339700        79625        49834
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━