Follow-up to the merged issue-74 work, applying an adversarial review of
commit 54b4d0f. Fixes two flake windows in the timing design, moves the
invariant checks to the right altitude, and clears the cleanup findings.

Correctness and flakiness:

- Reworked both reused-collector property tests (sampling integrity and
  conservation round-trip) to accumulate sent identities instead of
  resetting the sink between draws. A reset races an in-flight export
  from the previous draw; a late span could land after the reset and be
  misattributed to the next draw, failing CheckNoFabrication as a
  load-dependent flake. Over a running set, a late span is still a span
  that was sent.
- TestSampling_DeterministicDecisions now runs each of its two identical
  passes against its own collector and sink, so a late export from the
  first pass cannot bleed into the second. It also compares full span-id
  sets (premise and keep/drop), not just counts.
- Raised tailSettleIdle from 2s to 4s: the tail sampler's release
  latency is decision_wait (1s) plus its internal policy tick (~1s), so
  2s left no margin for scheduler jitter and could truncate the capture.
- waitForSpans now fails the test on a WaitFor timeout instead of
  silently returning a short read, so a hang is distinguishable from
  span loss.

Altitude and reuse:

- Moved the invariant checks into pkg/pipelinetest as exported,
  framework-agnostic functions returning errors: CheckNoFabrication,
  CheckParentsKept, CheckWholeTraces, plus CheckConservation for the
  round-trip tests, keyed by SpanKey and a Sent accumulator. External
  users can now compose their own pipeline tests on the harness, which
  the design doc promised; the always-on unit coverage moved with them
  to invariants_test.go. The synth test files now assert through this
  API. CheckWholeTraces derives trace grouping straight from received
  spans rather than re-parsing composite keys, and no caller rebuilds a
  received-key map the check also builds.
- Added pipelinetest.TracesConfig, a shared config skeleton parameterised
  by the processors block. The four hand-copied collector configs
  (pass-through, and the three samplers) now derive from it.
- Extracted randomTraceID/randomSpanID in replay.go over a next func()
  uint64 source; the replay generator and the test's seeded generator
  share the validity-retry loop instead of duplicating it.

Robustness:

- SupportsComponent now parses the collector's `components` output as
  YAML and matches component names only (not module paths), caching the
  set per binary path instead of spawning the binary and recompiling a
  regexp on every call.

Docs: the how-to's worked example now uses the exported pipelinetest API
and synth.GenerateTraces, so it is composable outside package synth;
config-building and tail-timing notes updated to match.

Co-Authored-By: Claude <[email protected]>
Claude-Session: https://claude.ai/code/session_017mb8JHoaMXNfA6f7sncP4A

Address review feedback on the sampling integrity harness:

- randomTraceID/randomSpanID now lay out their bytes with a fixed
  BigEndian order instead of NativeEndian, so a seeded ID source yields
  the same identities on any architecture.
- componentNames releases the cache mutex while it runs the collector
  `components` subprocess and unmarshals its output, re-taking the lock
  only to store, so a slow probe no longer blocks concurrent callers.
  The shared set is documented read-only.
- CheckNoFabrication and CheckConservation doc comments now state that
  they compare identity sets, not exactly-once counts: OTLP delivery is
  at-least-once, so a correct pipeline may redeliver a span on retry.
  Treating redelivery as a violation would flake against real
  collectors, so the set-based semantics are deliberate; only a missing
  or fabricated identity fails the check.

Co-Authored-By: Claude <[email protected]>
Claude-Session: https://claude.ai/code/session_017mb8JHoaMXNfA6f7sncP4A