feat(planning): add bounded semantic planner

[HayronHgh]

Summary

Implements 0.2.0-dev.4 — M4 Bounded Semantic Proposal Generation.

• build a deterministic, globally bounded PlannerInventoryView from Runtime-owned inventory facts
• add an isolated, tool-free Qwen planner using the versioned planner-v1 prompt
• disable Qwen thinking for the planner request while accepting only final message.content
• pass raw output through the frozen strict M2 parser, plan-ID challenge, inventory binding, and visible-unit binding
• invoke the unchanged M3 Runtime Validator and stop at ValidatedPlan
• add bounded retry/fallback behavior, session-only audit events, PAR/IPR and proposal telemetry
• add bilingual documentation, fixtures, fake-client tests, and a real llama.cpp/Qwen smoke test

Freeze-gate observability corrections

• result-level input/output tokens and latency are cumulative across all Planner attempts
• attempt audit events retain per-request token values
• parseFailures counts only malformed JSON, schema violations, and duplicate decisions
• failures are separated into protocol, binding, visibility, client, and stale categories
• planner-v1 now states the complete importance enum, non-empty reason rule, and positive targetTokens rule
• fake-client regressions cover cumulative retry cost and each failure class

Safety boundary

This PR intentionally does not add:

• context mutation or lifecycle/authority mutation
• Transformer or Executor
• artifact creation or memory promotion
• actual token-savings claims
• changes to ContextManager, AgentRuntime, or the M3 Validator policy
• validator-aware autonomous replanning

Hidden units remain implicit KEEP. Stale inventories are discarded without retry. Validator rejection stops and requires deterministic fallback.

Validation

• npm test: 111 passed, 1 Windows symlink-permission skip
• targeted planner/client tests: 16 passed
• JavaScript syntax checks: 36 files
• Markdown link checks: 28 files
• git diff --check: clean
• frozen boundary diff against main: no changes to ContextManager, AgentRuntime, MemoryStore, tool evidence, or Validator
• real local llama.cpp b10295 + Qwen3.6 E2E: V020_M4_E2E_OK

With the complete schema wording, the real Qwen smoke test passed the strict protocol on its first attempt (parseAttempts: 1, zero categorized failures). The unchanged M3 Validator then rejected one non-recoverable proposal and stopped with deterministic fallback and no mutation.

Follow-up invariant

M5 must revalidate recovery-source existence and integrity immediately before any future transformation or execution.