feat(execution): add validated context transformation

[HayronHgh]

Summary

Completes the validated transformation, atomic execution, and post-commit finalization chain for 0.2.0-dev.5:

ValidatedPlan
  -> D0-D2 Execution Preflight
  -> ExecutablePlan
  -> D3 Transformer
  -> TransformationCandidate
  -> D4 Post-transform Validator
  -> ValidatedTransformation
  -> D5 Atomic Executor
  -> ExecutionResult(COMMITTED)
  -> D6 Post-commit Finalization
  -> Inventory Rebuild
  -> Canonical Signed Accounting
  -> ExecutionReport
  -> STOP

D0-D4 remain zero-mutation. D5 is the sole deterministic message-context commit boundary. D6 is model-free derived observation and has no rollback authority.

Milestone progress

• D0 — execution protocol and frozen M4 experiment identity
• D1 — current-source RecoveryVerifier
• D2 — ValidatedPlan -> ExecutablePlan preflight
• D3 — immutable zero-mutation TransformationCandidate generation
• D4 — deterministic and semantic post-transform validation
• D5 — model-free atomic context executor
• D6 — inventory rebuild, canonical re-estimation, and signed reduction accounting

D6 finalization boundary

• D5 captures canonical pre-commit ContextManager.estimateComponents() output
• The committed result binds the exact tool-schema digest and M3 potential upper bound
• D6 accepts only a strict COMMITTED ExecutionResult
• Current contextGeneration must equal D5's committed generation
• The existing Context Inventory registry must still match inventoryBefore
• ContextInventory.synchronize() rebuilds from committed messages without replacing the registry
• Removed units become inactive; replacements retain stable IDs and refresh content/token cost
• Inventory ID/fingerprint reflects the committed context
• Before/after use the same ContextManager estimator, tool envelope, tool choice, and fixed overhead
• actualReductionTokens = before.totalTokens - after.totalTokens
• Negative reduction is preserved and never clamped
• Potential upper bound and observed actual reduction remain separate metrics
• Success emits an immutable ExecutionReport
• Drift/rebuild/accounting failure emits EXECUTION_FINALIZATION_FAILED with no actual-reduction claim
• D6 failure never becomes EXECUTION_ABORTED, relabels D5, or rolls back committed messages

Preserved invariants

• ValidatedPlan != ExecutablePlan
• M3 recovery classification is not execution-time recovery proof
• COMPRESS authorization does not authorize arbitrary replacement content
• ValidatedTransformation alone is not an execution capability
• D5 and D6 are model-free
• M4 planner-v1, inventory selection, budgets, protocols, policy, PAR/IPR, and telemetry remain hash-frozen
• No artifact/project-memory/episode write or semantic-policy change
• The branch stops at ExecutionReport -> STOP; no D7 scope

Accounting note

“Actual” means the observed signed difference under ContextOS's frozen-compatible canonical ContextManager estimator. It includes message serialization overhead, tool schemas/tool choice, and fixed prompt overhead; it does not claim tokenizer-exact llama.cpp backend counts.

Validation

• npm.cmd test: 172 tests, 171 passed, 1 Windows symlink capability skip
• D5-D6 focused tests: 25 tests
• D6 coverage includes stable-ID rebuild, inactive removed units, fingerprint change, canonical before/after envelope, identical tools/overhead, signed negative reduction, generation drift, failure non-rollback, immutable report, and end-to-end execution
• JavaScript syntax: 57 files
• JSON parsing: 19 files
• Markdown local links: 110 links across 30 files
• git diff --check
• M4 frozen-input diff against aa59f4d: empty
• D6 model/external-write scan: clean
• Signed reduction clamp scan: clean

Documentation

Updated English and Traditional Chinese README, changelog, architecture, execution contract, technical report, and RFC.

This PR remains Draft as the complete dev.5 D0-D6 freeze candidate.