[feat] Read the sessions of more tools, so they can be scored #1418

Why

teamai scores a session only when it can read the tool's transcript. Today that is Claude Code, Codex, Cursor, CodeBuddy, WorkBuddy and OpenCode (plus the internal builds). The other 14 agents in the README matrix show — under scorers, and no open issue adds them: #1385, #1386 and #1388 all leave "reading sessions of tools without an adapter" out of scope. #1386 lets any driver's tool judge; this issue lets teamai read more tools' sessions. Together they make scorers work for as many tools as possible.

 TOOLS[tool].sessions
-  qoder, qoder-cn, pi, omp, copilot, muse, hermes, openclaw   'none'
+  qoder, qoder-cn, pi, omp, copilot, muse, hermes, openclaw   ['<platform>']
 README matrix, scorers column
-  ✓ 6 agents
+  ✓ 14 agents

What each tool keeps

Evidence: v seen on a real install, s the tool's source or shipped bundle, d docs only. A third-party parser is a lead, not evidence.

Behavior

• An adapter for these tools may be read-only. Scoring needs listConversations, readSession and existingSessionId; writeSession and deleteSession throw ToolUnavailableError naming the tool, as Cursor and OpenCode already do on paths they cannot write. teamai session migrate --target into such a tool fails with that message; nothing else changes for it. Qoder writes Claude Code's format, so it keeps the writer ClaudeCodeAdapter already has.
• Each tool lands in its own PR, in the order of the table. The issue closes when every ready and needs work row is done; the deferred, unknown and blocked rows stay 'none' or 'unknown' in TOOLS with a note saying why.

Acceptance criteria

1. For each adapter added: teamai session platforms lists the tool, and teamai session push --source <platform> --scrub archives its sessions for the current project with title, cwd, timestamps, user prompts, assistant replies and tool calls in order.
2. teamai score --scorer <s> --session <id> --source <platform> labels a session of that tool, judged by an installed judge tool.
3. The Pi-family adapter reads both Pi's and OMP's files, including OMP's title line and its folder names, and takes the cwd from the session header.
4. Hermes and OpenClaw list only coding sessions, not chat-platform ones.
5. teamai session migrate --target <platform> into a read-only adapter fails with the ToolUnavailableError message and writes nothing.

Definition of done

• TOOLS[tool].sessions names the new platform. The README scorers tick, in every README*.md, follows: readme-matrix.test.ts fails until it does.
• docs/guide/tool-support.md and its Chinese page regenerated with npx vitest run tool-support -u.
• The notes in src/tools.ts match what is now known: Pi's version (1.4.0) and sessionDir setting, an Oh My Pi note, QODER_CLI_HOME for Qoder, Muse's sessions under XDG_DATA_HOME.
• docs/guide/scorers.md and its Chinese page list the tools whose sessions can be scored, if they list them.
• Real-CLI record per PR: teamai session push --source <platform> and one teamai score --session <id> --source <platform> on a real session of the tool.

Out of scope

• Which tool judges ([feat] Judge a session with the driver's own tool #1386).
• Decrypting Trae's IDE database.
• A zstd decoder for Node 20. OpenClaw and DeepSeek Harness wait for one, or for Node 22 as the floor.
• Writers for the new formats.

Dependencies

• Needs feat(sessions): resolve each tool's session store through locate (#1221) #1404 merged: it adds the sessions facet to TOOLS ([feat] One descriptor per tool, and one resolver for where teamai writes and where the tool reads #1221 F).
• Independent of [feat] Judge a session with the driver's own tool #1386. Once both land, a session from any of these tools can also be judged by the driver's own tool.