Examples validation harness: scripts/validate_examples.py + manifest + CI tiers

[Pendu]

Part of the 0.3.7 testing program (#322). Prerequisite: #324
(step knobs). Absorbs the CI-execution side of #319.

Problem

The 17 examples are the de-facto integration suite: they are the only runnable
artifacts that exercise the full launcher -> aggregator -> training pipeline
the way users do. No CI job runs any of them. During the 0.3.6 pre-merge
validation, a scripted sweep of the CPU-runnable examples caught more product
surface per minute than any test directory: verdict flips, per-rank rows,
null-vs-zero semantics, manifest lifecycle, stderr cleanliness.

Proposal

1. scripts/validate_examples.py: runs manifest-listed examples sequentially
   under the real launcher, one TRACEML_LOGS_DIR per example, and asserts per
   run: exit 0; final_summary.json present with the current schema version;
   run manifest completed; h2d_ms null (never 0.0) on CPU; the expected
   diagnosis assertion (see tiers below); no [TraceML]-prefixed
   error/traceback lines in stderr.
2. A manifest (checked in, single source of truth) where each entry declares:
   script, args, hardware (cpu/gpu/multi-gpu), timeout (~300s cap), and an
   assertion tier:
   • structural: rc/schema/manifest/nulls only (default for anything
     timing-sensitive on shared runners),
   • family: verdict must be in an allowlist (only where an injected knob
     gives a large signal margin, e.g. the bottleneck demo's slow scenario),
   • exact: reserved for structurally-forced outcomes (e.g. a
     no-instrumentation example must produce INCOMPLETE_DATA).
3. CI wiring per the umbrella tier model: CPU subset on PRs targeting
   version_* branches (paths-filtered) + nightly full sweep; NOT per-PR
   (except the Add a 2-process gloo DDP end-to-end test to CI #319 DDP entry, which is this harness invoked on one manifest
   entry). Job hardening: OMP_NUM_THREADS=1, job timeout-minutes, file-based
   or port-0 rendezvous for the DDP entry, artifacts (logs dir + summary +
   stdout/stderr) uploaded on failure, one-line contributor-legible failure
   messages (example, assertion class, expected vs actual), and a post-run
   no-orphan-process check.
4. The examples/README expected-verdict table is GENERATED from the manifest
   (or CI asserts they agree); two hand-maintained copies of the same truth
   is how fixtures drifted before.

Notes

• All current CPU-runnable examples are synthetic and download-free; the
  manifest rule "no network in CI entries" keeps it that way.
• The maintainers' release-validation checklist invokes this same script;
  no forked sweep logic (engine invariant in the umbrella).