Pack promotion misses natural-language server-name mentions (e.g. "claude-projects mcp" → workspace-critical not promoted)

[samscarrow]

Symptom

Prompt mentioning an MCP server by name (e.g. using the claude-projects mcp, …) does not promote the pack containing that server. Tools remain deferred and invisible to the model. The user only recovers by typing a fully-qualified tool name like mcp__claude-projects__claude_list_projects, which trips the explicit-mention rescue path.

Evidence (from ~/.tcp-shadow/proxy/decisions.jsonl, 2026-05-04 post-IMP-21)

Failed turn — 19:44:54 UTC

prompt: "using the claude-projects mcp, Replace updated files on claude.ai pontius project and upload new ones too"
tool_count_before/after: 5 -> 5
active_packs:    ["core-coding"]
deferred_packs:  ["workspace-critical"]   ← contains claude-projects
workspace_allowed_servers: ["bay-view-graph"]
claude-projects survivors: 0

Recovered turn — 19:45:38 UTC (user pasted explicit tool name)

prompt: "...e.g.: mcp__claude-projects__claude_list_projects"
tool_count_before/after: 293 -> 188
survivors include: mcp__claude-projects__claude_list_projects, claude_list_docs, claude_read_chat, ...

The recovery is the explicit-mention rescue, not a derivation win.

Diagnosis

Pack promotion appears to key on tool-name fragments / capability flags, not on server-name mentions in natural prose. "claude-projects" is the server identifier, not a tool keyword, so the workspace-critical pack stays deferred even though the user clearly named it.

Proposed direction (not pre-decided)

Add a server-name lexical signal to pack-promotion derivation: if a deferred pack contains server S and the prompt contains S (or a normalized form like S-mcp / S server / mcp__S__) as a token, promote the pack.

Edge cases to think through before implementing:

• Server names that collide with English words.
• Hyphen vs underscore normalization (claude-projects vs claude_projects vs mcp__claude-projects__*).
• Whether to promote on mention alone, or require a co-occurring action verb.
• Interaction with TCP-IMP-17 prompt-similarity ranking.

Relation to existing work

• Concrete instance of the bottleneck named in CLAUDE.md: "Request derivation precision (5.4% in TCP-VAL-1). Hardening this logic is the top priority for unblocking the validation layer."
• Should be measurable once TCP-MT-20 has enough scorable post-IMP-21 turns: this case would show as a strict miss with expected_tool_name not in tool_call_sequence because the tool was filtered out before the model could call it. Also note: expected_tool_name may itself be unset on filtered turns — telemetry interaction worth checking when this is repaired.

Repro

Issue any prompt of the form using the X mcp, do Y where server X is in a non-active pack. Inspect the resulting decisions.jsonl row and confirm X-pack stays in deferred_packs with zero survivors.