Preserve tools after skill activation

[supercobra]

Summary

• preserve the parent conversation's local tool surface when the model calls activate_skill
• continue registering tools declared by an activated skill without applying its allowed-tools as a persistent engine-wide filter
• retain scoped allowed-tools enforcement for direct and isolated skill invocations
• document the distinction between model-driven and direct skill activation

Reproduction

Added a regression test that activates a restricted skill with a large instruction payload, then verifies read_file, shell, and the skill-declared tool remain visible:

• on the immediate agentic follow-up after activation
• on a subsequent user turn (continue)

Before the fix, both post-activation requests exposed only the skill-declared tool.

Verification

• go test ./cmd -run 'Skill' -count=1
• go test ./internal/tools -run 'Skill' -count=1
• make build
• go vet ./...
• make complexity
• git diff --check

go test ./... was also run. It reached unrelated existing failures in internal/jobs when inheriting Fish as $SHELL (those tests pass with SHELL=/bin/sh) and a timestamp-ordering failure in internal/session/TestSQLiteStoreUpdateMessageMovesSessionByLatestResponseActivity.