fix(skills): preserve scoped models across async continuations

[guustavomv]

Motivation

Custom planning skills can lose their declared model after a background-agent completion notification: an initial /plan request begins on gpt-5.6-sol, while a queued continuation falls back to the session model. A separate production failure affected OpenAI-compatible tool calling: the configured max/deepseek-v4-pro execution coordinator repeatedly emitted Rea instead of Read, burning calls before falling back to Bash.

Changes

• Persist skill effort in the inline command_permissions attachment.
• Recover the latest skill model, effort, and allowed tools for task-notification continuations.
• End inherited skill scope after a later visible user prompt.
• Assemble streamed function-name fragments such as Rea + d before tool finalization.
• Recover an unambiguous truncated built-in tool name, including after a filtered execution path omits its original definition. MCP names and ambiguous/short prefixes are never guessed.
• Apply the same recovery to streaming and non-streaming tool execution paths.
• Add regressions for model-scope continuation, split tool names, and safe prefix recovery.

Reproduction evidence

From an interactive /execute session with --debug-file:

• The session UI displayed max/minimax-m3, but the effective skill API calls were provider: openai, model: max/deepseek-v4-pro.
• Before the fix, debug logged repeated Unknown tool Rea entries, and the transcript contained repeated Error: No such tool available: Rea results.
• The model then abandoned Read and used Bash, which succeeded. This isolated the failure to tool-name handling rather than repository access or permissions.

The companion orchestration-pack mitigation is published separately in guustavomv/verboo-orchestration-pack commit 5bc7e20: /execute no longer exposes Read and instructs the coordinator to read contracts using read-only Bash commands. A later /execute run made only Bash calls and showed no new Unknown tool Rea errors.

Planning bundles intentionally remain outside the harness configuration directory at /.plans/ through orchestration-pack commit f726c17; no special permission exception is included in this PR.

Validation

• bun test src/Tool.test.ts src/services/tools/StreamingToolExecutor.toolPairing.test.ts src/services/api/openaiShim.test.ts src/utils/processUserInput/activeSkillScope.test.ts: 122 passed, 0 failed.
• bun run build: CLI and SDK bundles built and validated.
• GitHub PR checks: smoke, unit, provider, Python, web, and intent checks passed for the prior head; the current revert commit is pushed and awaiting its PR checks.

The repository-wide tsc --noEmit still reports the existing upstream baseline of unrelated missing modules/type errors; the touched production paths introduce no new targeted errors.