Cleanup batch + harmonic lib tests + ROADMAP v2

Final session push: knock out cleanup-tier items, add real test
coverage for the harmonic libraries via OMC's own --test mode,
publish ROADMAP v2 reflecting current state.

== Cleanup ==

* Float display always shows decimal point. `println(3.0)` now
  prints "3.0" instead of "3". Fixes PAIN_POINTS LOW-1; removes
  int-vs-float ambiguity from user output. Centralised via
  format_float helper in value.rs; println / print_raw routed
  through to_display_string instead of inlining their own match.

* `omc --audit FILE` mode runs FILE under both engines (tree-walk
  and VM via OMC_VM=1) and exits 1 if stdout differs. Fixes
  PAIN_POINTS LOW-2. Catches the class of bug the float-truncation
  issue surfaced as: silent engine divergence with no signal.
  Spawns child processes via std::process::Command — uses the
  same binary the user would invoke directly, so audit mirrors
  real execution exactly.

== OMC test suite ==

* examples/tests/test_harmonic_libs.omc — 18 tests covering
  harmonic_anomaly, harmonic_clustering, harmonic_recommend, plus
  regression tests for the equality bug fix from the previous commit
  (dict / function / array / string / 0 / "" all explicitly tested
  != null). Run via `omnimcode-standalone --test
  examples/tests/test_harmonic_libs.omc`.
  All 18 pass.

  Eats own dogfood: tests for OMC libraries are themselves OMC
  programs running through OMC's --test harness.

== ROADMAP v2 ==

* Marked supersedes: ROADMAP.json v1.0
* shipped phase: added 9 items from this session (elif, --test,
  --bench, --audit, harmonic_clustering, harmonic_recommend,
  attack_zoo, float_display_fix, equality_bug_fix, harmonic-lib-tests)
* next-1-3-sessions: rotated in 6 new items (csv_parse builtin,
  selective imports, heal-pass context awareness, harmonic_anomaly
  on real telemetry, VM-native match, incremental fit for online
  learning)
* New session_summary_2026_05_15 block with consolidated wins +
  honest negative results
* New lessons-learned phase with 6 meta-insights extracted from
  this arc (real data finds bugs, honest negatives, eat own dogfood,
  architecture thresholds, substrate not syntax, integration in
  host language)
* principles + anti_goals refreshed; added "eat your own dogfood"
  and "cherry-picking results is forbidden"

43/43 functional examples produce identical output under tree-walk
and VM. 18/18 OMC tests pass via --test mode. 92/92 unit tests pass.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: RandomCoder-lab