fix(acp): spell out every bb skill chip the agent did not advertise

[SawyerHood]

Human comments

What was wrong

A bb skill chip reaches an ACP agent as literal /name text. OpenCode reads a leading /name as one of its own commands and ends the turn without calling the model (#4923).

The rebuilt bridge rewrote a chip only in narrow cases:

• only a chip at the very start of the first text item,
• only once the agent had advertised its commands.

A chip on the first turn, mid-message, in a later text item, or next to an advertised command still went through raw.

What changed

textWithBbSkillCommandsSpelledOut (bridge.ts) replaces textWithoutUnknownLeadingSkillCommand:

• What it rewrites: every slash chip whose name is a configured bb skill (skills/configure roots), in every text item.
• How: to Use the bb skill "name", with a trailing colon when the chip leads further text. Chips are processed right to left, skipping overlapping or mismatched ranges.
• What it leaves alone: the agent's own advertised commands, chips that are not bb skills, and unmarked slash text.

This follows #5055 by @timurkhakhalev (credited as co-author), rebuilt on the new bridge.

How you verified

• The rewritten bridge test covers six turns:
  • first-turn chip,
  • two chips mid-message,
  • a non-bb chip next to a bb chip in a second text item,
  • a mismatched range,
  • after the agent advertises commands,
  • an advertised command next to a bb chip.
• Under the previous rule, the first-turn, mid-message and mixed cases come out raw.
• pnpm exec turbo run test lint typecheck --filter=@bb/provider-bridge-acp --filter=@bb/provider-parity: bridge 522 passed, parity 59 passed.
• Live with opencode 1.18.35: a message with a project skill chip in the middle ("First, follow /e2e-greet …") got the skill's exact reply.

Fixes #4923. Supersedes #5055.

AGENT GENERATED