docs(commands): teach Claude to extract PR refs and quote focus text

[JohnnyVicious]

Summary

Fixes the bug you spotted: `/opencode:adversarial-review on PR #389` was silently misrouted to a working-tree review because bash strips unquoted `#389` tokens as comments before they reach the companion script.

`detectPrReference` (added in #2) was correct, but it never ran — by the time argv reached node, the `#389` had already been chopped off at the shell boundary, so the regex had nothing to match. Demonstration:

```
$ bash -c 'node -e "console.log(JSON.stringify(process.argv.slice(1)))" adversarial-review on PR #389'
["adversarial-review","on","PR"] ← unquoted: #389 stripped

$ bash -c "node -e "console.log(JSON.stringify(process.argv.slice(1)))" adversarial-review 'on PR #389'"
["adversarial-review","on PR #389"] ← single-quoted: preserved
```

The fix has to live at the slash-command markdown level, because that's where Claude reads the instructions about how to construct the bash command. Code-level fixes are powerless once bash has already eaten the input.

Changes

1. PR reference extraction (REQUIRED — both review.md and adversarial-review.md):

   • New rule telling Claude to recognise PR #N / pr #N / PR N patterns in the user's input, extract the number, pass it as --pr N, and strip the matched phrase from any focus text.
   • Includes a worked transformation example: on PR #390 challenge the caching strategy → --pr 390 'challenge the caching strategy'.

2. Focus text quoting (REQUIRED — adversarial-review.md only, since review.md doesn't take focus text):

   • New rule telling Claude to wrap focus text in single quotes when invoking bash, so other shell metacharacters (*, $, ;, &, |, <, >, parens, backticks) survive intact.
   • Includes the standard '\'' escape recipe for apostrophes in user input.

3. Foreground/background flow examples updated in adversarial-review.md:

   • Stops showing $ARGUMENTS literally (which was training Claude to invoke bash without any of the above transformations).
   • Instead shows the illustrative shape <flags> [--pr N] ['<quoted focus text>'] with an explicit "transform $ARGUMENTS first" instruction pointing at the rules above.

Companion-script behaviour kept as fallback

detectPrReference and the auto-detect path in handleAdversarialReview are intentionally kept, so that:

• Anyone calling the companion script directly via bash with proper quoting still gets PR detection
• Future invocation paths that don't go through the slash command still benefit
• The defensive fallback covers Claude not following the new rules perfectly

Test plan

• Local: npm test — 62 tests, 0 failures (markdown-only changes don't affect tests)
• Demonstrated the bash comment-stripping behaviour with a minimal repro
• Manual verification after merge: /opencode:adversarial-review on PR #389 should produce node ... adversarial-review --pr 389 and the review should hit PR #389's diff, not the local working tree
• CI green