fix(examples): an example built WITHOUT its feature is not a broken example — needs-feature

[noahgift]

examples-nightly is one of five red scheduled lanes. Its verdict is currently not a
trustworthy signal about whether examples work — which is the exact property
APR-RELEASE-001 §4.2 and T-1 want to depend on.

Run 34826552378, 980 example targets

summary pass=602 fail=22 timeout=6 needs-args=30 needs-hardware=2 needs-data=318

The 28 fail+timeout rows, classified by what the output actually says:

 6   the example is not built with the feature it needs   <- NOT broken
 2   needs a network peer (192.168.50.100:9000)           <- NOT broken
 1   needs a TTY ("No such device or address")            <- NOT broken
 6   timeout at 180 s (buggy_server, brick_computer ...)  <- servers by design?
13   remainder

A quarter of the lane's red is a missing label, not a defect. This PR takes the 6:

aprender-db::compressed_kv          This example requires the 'compression' feature.
aprender-rag::compressed_index      This example requires the 'compression' feature.
aprender-rag::semantic_embeddings   This example requires the 'embeddings' feature.
aprender-distribute::tensor_example Error: This example requires the 'tensor' feature.

plus cargo's own refusal, which phrases it differently:

target `gpu_info` in package `aprender-zram-core` requires the features: cuda

Like needs-hardware, not like needs-args

A needs-feature row is a SKIP that must be re-run WITH the feature before a release
verdict. A green lane on a default-feature host must never be readable as "the compression
example works" — which is why the row cites the line that classified it.

This follows the precedent the file set for needs-data on 2026-09-12 (173 rows), and obeys
the rule its own header states: "one turns a real defect into a skip, so each addition belongs
in the selftest."

The regex ships both polarities

A skip class can only ever turn a fail into a non-failure, so the dangerous direction is
tested explicitly:

MUST match      This example requires the 'compression' feature.
                Error: This example requires the 'tensor' feature.
                ... requires the features: cuda
MUST NOT match  thread 'main' panicked at src/main.rs:11:39
                error[E0432]: unresolved import `trueno_gpu::driver::PinnedBuffer`
                Error: Os { code: 6, kind: Uncategorized }
                assertion failed: left == right
                feature is required for this operation     <- the deciding near-miss

Mutation, both directions:

regex matching nothing       -> selftest RED
regex widened to "feature"   -> selftest RED
shipped regex                -> green

The harness caught its own staleness: a completeness assertion hardcoded "one of the six
classes" and went red on the seventh. Updated rather than relaxed — that assertion is what
stops a row being written with no class at all.

Deliberately not swept in

The 2 network rows, the 1 TTY row, the 6 timeouts and the 13 remaining are separate causes and
deserve their own triage. One of the 13, aprender-zram-core::gpu_info, is already owned by
#3267, which found that trueno_gpu::driver::PinnedBuffer "does not exist and never did in
this monorepo".

bashrs 7.4.1: 0 errors. selftest: all rows passed.

no-close: the measurement this implements is #3182 (open, 0.68.0), which counted 9
of 44 non-pass rows as feature refusals miscalled fail. This PR adds the needs-feature
class for exactly those rows; #3182 also covers the 20 long-running-by-design timeout
rows and the needs-data surface, so it stays open. The four other red nightlies
(Coverage Nightly, Silicon Nightly, CUDA Nightly, and the rest of examples-nightly)
are separate causes.

ont-delta: none — a classification fix in a CI harness; it adds no entity, shape, verdict
reason or resolves target of its own.

🤖 Generated with Claude Code

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=3288 head=106194ccc91107196650dc8c650323a1d09b9914 verdict=REFUSE class=Q1 arm_rc=1

Shadow mode: this records a verdict and merges nothing. A refusal
to arm is not a block (§13 adds zero rows to §7) — the pull request is
exactly as green as it was.

[noahgift]

Superseded by #3295, which carries this change unmodified along with eight other
collision-free CI/guard PRs as a single merge-queue entry.

Not abandoned and not rejected: the commits are merged into
PMAT-3231-batch-1-ci-guards verbatim (clean=9 conflicted=0), and this branch is
retained — reopening is one click if #3295 does not land.

Why: a successful merge_group CI run is median 78 min with max_entries_to_merge = 1,
so nine entries is ~12 h of queue for 623 insertions. As one entry it is 78 minutes.
Root cause and the full five-whys: #3294.