test: make the default suite reproducible in clean environments

[Marsu6996]

Draft / stacked PR

Do not merge before #67. This branch is stacked on #67, so the GitHub diff
includes that dependency until it lands. It will then be rebased onto main
to leave only the baseline-reproducibility changes.

Problem

The default suite contained assumptions tied to a particular checkout, runner
load, or process history: an undeclared XML parser, a local planning artifact,
global stderr monkeypatching, hard-coded package versions, cache-key
expectations that no longer matched the write contract, platform-sensitive
vector normalization, and an assertion comparing asynchronously sampled macOS
memory metrics.

Two concurrency/ranking assertions also over-specified wall-clock behaviour:

• the GIL probe appended roughly 23–24 million Python floats while scanning,
  approaching 1 GiB of probe-owned allocations and interpreting its own list
  reallocations as GIL stalls;
• the empty-profile fallback test required identical ordering for records whose
  scores differed only by per-candidate age-sampling noise.

Solution

• use the standard library XML parser for launchd plist checks;
• measure community detection directly instead of requiring a planning file;
• inject the warning writer instead of mutating global stderr;
• derive the package version from source metadata or installed distribution
  metadata, with a clear source-only fallback;
• align cache and normalization tests with their production paths;
• keep the community performance hard cap while making the historical soft
  target an observable warning;
• remove the unstable phys_footprint <= RSS comparison while retaining the
  deterministic watchdog decision tests;
• keep LongMemEval preflight tests hermetic by stubbing the cleaned adapter's
  network-resolved revision together with its dataset loader;
• count only iai_mcp_native extension binaries in the native import test,
  rather than every unrelated extension installed in site-packages;
• replace the wall-clock GIL ceiling with a synchronized observer: periodic
  Python switching is temporarily disabled and the observer must make progress
  inside py.allow_threads(...) during the native scan;
• compare empty/medium fallback membership and score by record ID, preserving
  the existing tolerance without assigning meaning to quasi-tie order;
• register the existing literature marker.

Guarantees and scope

No production safety gate is relaxed. The patch removes machine-specific test
assumptions and keeps the correctness signals: the GIL test still fails against
an intentionally GIL-holding C call, and the profile test still catches any
record-set or per-record score divergence.

Validation

• synchronized GIL probe: 20/20 local repetitions; 10/10 under 20 concurrent
  CPU workers; negative control 0/3 progress with an intentional GIL holder;
• Rust GIL/Send suite: 6/6;
• empty/medium fallback: 50/50 repeated runs; full knob file 9/9;
• focused final-stack Python set: 40 passed, 1 skipped;
• wrapper suite: 48/48, plus TypeScript typecheck;
• exhaustive default suite on the prior stack tip: 5,159 passed, 212 skipped,
  4 xfailed, 0 failed across 5,375 collected tests; GitHub's single-process
  rerun is the final publication gate for this amended tip.

Designed by Marsu — Refined by Codex.

Public CI result

The amended macOS single-process run completed with 5,123 passed, 113 skipped, 107 deselected, and 4 xfailed. Its only three failures are the pre-existing src versus IAI-MCP spatial-classification cases fixed by the next stacked PR, #69.
The prior daemon pre-warm, GIL-concurrency, and quasi-tie ordering failures are absent from this run. The cumulative #69 tip is fully green at 5,126 passed and also builds the universal2 wheel successfully.