v0.1.2 durability and core invariant freeze

[HayronHgh]

What changed

• split tool-evidence persistence from prompt rendering
• persist medium and large tool results with character/byte counts and SHA-256
• add machine-checkable internal context_os recovery metadata
• add an internal-to-model serialization boundary so runtime metadata never enters the OpenAI-compatible protocol
• gate 55% stale-output compression and 65% complete-exchange eviction on artifact recoverability
• preserve artifact recovery references in markers, State Transfer, rebuilt prompts, and working-state reads
• add the twelfth runtime tool, read_artifact, with ID-only, 500-line-bounded, integrity-checked retrieval
• define episode retrieval as latest N valid episodes and recover corrupted repository maps as cache misses
• add durability observability counters
• freeze the deterministic Phase 1/2 architecture and reserve adaptive semantic planning for v0.2.0
• update English and Traditional Chinese documentation and package/runtime version to 0.1.2

Why

Before this change, a medium tool result could exist only in active context and later be shortened or removed without an exact recovery artifact. Natural-language prune markers were not machine-verifiable. Corrupted newest episodes could also hide older valid memory.

v0.1.2 makes deterministic context reduction recoverable and fail-loud before Phase 1/2 is frozen.

User and developer impact

• default durability threshold: 800 characters
• default active prompt limit: 12,000 characters
• invalid threshold ordering is rejected at startup
• tool count increases from 11 to 12
• existing local configs remain compatible through normalized defaults
• /compact and threshold keys retain v0.1.1 behavior

Validation

• npm.cmd test: 35 tests defined, 34 passed locally, 1 file-symlink test conditionally skipped because this Windows host denies file symlink creation
• Windows junction file/artifact escape tests passed
• all source files passed node --check
• 3 JSON files parsed
• 5 PowerShell scripts parsed
• 20 Markdown files passed relative-link validation
• sensitive-information and staged-diff checks passed
• real llama.cpp + Qwen3.6 E2E passed: read_file -> 14,116-character artifact -> read_artifact -> V012_E2E_OK