fix(pi-kilocode): security hardening

[iRonin]

Summary

Several defensive fixes for scenarios where the Kilo Code API—or a
compromised network path—returns unexpected data.

Changes

src/lib/env.ts

• Guard PI_CODING_AGENT_DIR against null-byte injection before passing
  it to path.resolve()

src/provider/oauth.ts

• Cap auth responses at 1 MB (checked via content-length header and
  actual body length) to prevent memory exhaustion
• Change fetchJson return type from T to T | null; all three
  call-sites (initiateDeviceAuth, pollDeviceAuth, fetchProfile) now
  explicitly guard against a missing or unparseable body instead of
  receiving undefined silently cast as T
• Reject org IDs containing control characters (CR, LF, NUL, etc.)
  before injecting them into HTTP headers — defence-in-depth alongside
  the protections already in modern fetch implementations

src/provider/models.ts

• Add isValidCacheFile() type guard; readCache() now validates the
  JSON structure before trusting it, so a tampered local cache file
  cannot inject arbitrary model configuration
• Cap model-list fetch at 50 MB and assert the data array shape
  before returning

Tests

• 14 new tests covering each new guard: response size limits, invalid
  JSON, header injection via org ID, cache file tampering

package-lock.json

• npm audit fix — resolves 4 moderate advisories (brace-expansion,
  fast-xml-parser, yaml, @aws-sdk/xml-builder)

[sudosubin]

Thanks for the change. After reviewing the threat model, I found that these guards address scenarios that are highly unlikely to accour in this provider, so their utility does not appear to outweigh the complexity of the code. Therefore I'll close this for now.