fix: review findings (MD037 underscores, bun.lock consistency)

- PR template line 47: wrap latency placeholders in backticks so MD037 does not
  parse `___` as emphasis.
- frontend/CLAUDE.md off-limits section: change `bun.lockb` to `bun.lock` to
  match the actual lockfile on disk and the rest of this file. Reviewer
  suggested the reverse, but verification against the working tree shows the
  text-format `bun.lock` is what Bun is using here.

Skipped review finding: convert bold labels under "What changed" to ATX
headings. Those labels are inline bold text, not Setext headings, so MD003
does not apply. Adding `###` headers per stack would add visual weight that
contradicts the "lighter, less generic" direction this template was built for.

Note: the root CLAUDE.md also has a `bun.lockb` reference inconsistency
(off-limits section). Leaving that for a separate follow-up since it is not
part of this PR.

Author: DevanshuNEU

.github/PULL_REQUEST_TEMPLATE.md

@@ -44,7 +44,7 @@
 
 - [ ] Data model matches ADR
 - [ ] Consistency guarantee matches ADR (strong / read-after-write / eventual)
-- [ ] Latency target honored, measured: p50 ___ ms / p95 ___ ms (ADR target: ___)
+- [ ] Latency target honored, measured: p50 `<value>` ms / p95 `<value>` ms (ADR target: `<value>`)
 - [ ] Blast radius is what the ADR predicted
 - [ ] All ADR acceptance criteria pass
 

frontend/CLAUDE.md

@@ -78,6 +78,6 @@ Most files stay under 200 lines. These are approved larger due to logical cohesi
 
 ## Off-limits files in frontend/
 
-- `bun.lockb`
+- `bun.lock`
 
 Never delete or hand-edit. Let Bun manage it.