v0.1.12 — progressive_reads telemetry + timeout hardening

Wraps up two workstreams into a single release:

Issue #204 — trace_id + progressive_reads

• trace_id threaded through the progressive delivery path so Langfuse can correlate a stm_proxy_read_more span with its originating proxy_call turn (#205).
• New stm_progressive_stats MCP tool + progressive_reads table in ~/.memtomem/stm_feedback.db — reports total reads, distinct responses, follow-up rate, avg chars served, avg coverage, and a per-tool breakdown (#213).

MCP hang / timeout hardening

• Upstream proxy calls respect call_timeout_seconds + overall_deadline_seconds (#206).
• LLM compression stage respects compression.llm.llm_timeout_seconds (#207, #210).
• Bounded-lock helper + lock_timeout_seconds on internal ProxyManager locks (#208, #211).
• Exception barrier around mcp.run() so unhandled tool exceptions no longer leave the stdio subprocess half-dead (#209 Part A, #212).

Rolled in from Unreleased

• MEMTOMEM_STM_ADVERTISE_OBSERVABILITY_TOOLS=false hides the seven operator-facing tools from tools/list for eager-loading MCP clients (#201).
• mms add --from-clients / --import bulk-imports servers post-init (#200).
• mms add --import dual-registration warning (#202, #203).

Also syncs src/memtomem_stm/__init__.py __version__ back from its 0.1.9 drift.

Full detail: CHANGELOG.md