fix(deps): retarget four security overrides whose keys had gone inert

[choraria]

Clears all 8 open Dependabot alerts (2 high, 6 moderate). No Dependabot PR exists for these — every
one is a transitive dep governed by the overrides block, which Dependabot can't author a patch
for. The updater job ran and opened nothing.

Root cause: four pins had silently gone inert

Every one of these packages was already pinned. All four keys stopped matching when their
advisory moved the fix floor above the key's fixed — the exact failure mode the block's own CAUTION
comment describes:

key	tree sat at	new floor	matched?
undici@<7.28.0	exactly 7.28.0	7.29.0	✗ key can't match its own boundary
fast-uri@<3.1.3	3.1.4	3.1.5	✗ drifted above the ceiling
hono@<4.12.27	4.12.32	4.12.34	✗
postcss: 8.5.18 (bare)	8.5.18	8.5.23	✗ bare pin is a floor and a ceiling

pnpm reports nothing when a key goes inert, and overrides-guard only proves a pin moved the
tree — it explicitly disclaims advisory drift ("Dependabot's job, not this guard's"). So the tree
sat vulnerable behind a green guard. Retargeted, not widened; each key still touches only its
vulnerable band.

Reachability — re-measured, not carried forward

The bundle census in the comment block was 521 inputs; it's 648 now, so the old numbers couldn't
be assumed forward. Re-run via wrangler deploy -c apps/mcp/wrangler.jsonc --dry-run (3,295,870 bytes):

• fast-uri — 3 of 648, genuinely ships. The new advisory is host confusion via a backslash
  authority introducer. Reached only as ajv's uriResolver for schema $id/$ref — re-verified:
  the sole call sites are dist/compile/resolve.js and the dist/core.js default, and ajv-formats
  3.0.1 still implements format: "uri" as a regex and never calls it. Request payloads never reach
  fast-uri's host parser, so the HIGH rating (which is about gating on a parsed host) doesn't carry
  over. ^3.1.5 is in-range for ajv's ^3.0.1.
• hono — 0 of 648. Re-measured rather than inherited, because the CORS ReDoS is a different code
  path from the JSX advisories the earlier census covered. Resolves to 4.13.0, in-range for both the
  SDK's ^4.11.4 and @hono/node-server's ^4.12.8.
• undici — 0 of 648. Two copies, only one vulnerable: miniflare's 7.28.0 (in-band) and jsdom's
  8.9.0, which is already at the 8.x fix floor — all five advisories list >= 8.0.0, < 8.9.0 → 8.9.0. The <7.29.0 key deliberately leaves the 8.x copy alone.
• postcss — 0 of 648. Build-time only, first-party stylesheets.

The one pin that changes executed code

undici is the only one of the four that touches something we actually run, and it crosses an exact
declared version — miniflare pins undici: 7.28.0, not a range. A bad outcome there surfaces as a
test flake, not a build error, so it was verified rather than assumed: apps/engine (35 files / 460
tests) run 3x on each side — 460/460 passed all six times.

An EnvironmentTeardownError appears in the log nondeterministically. A single A/B sample made it
look like the bump caused it (2 occurrences with, 0 without) — it doesn't. Three runs per side:
7.28.0 emitted it in 2 of 3 runs, 7.29.0 in 0 of 3. Pre-existing race, unrelated to this change.

Test plan

Full gate, TURBO_FORCE=true throughout:

• pnpm lint — 997/997 · format:check clean · typecheck 33/33 · test 29/29 · build 24/24
• pnpm no-skipped-tests — clean
• pnpm overrides-guard — all 14 pins documented and moved; no copy below its floor across 1290 packages
• apps/engine 3x per side of the undici bump (see above)
• pre-push hook — full gate incl. db:test