fix(extension): avoid recapturing recalled memories

[GautamSharma99]

Summary

When auto-search and auto-capture were enabled together, recalled memories were appended to the submitted prompt and then persisted through
CAPTURE_PROMPT as if they were new user input.

This caused retrieved memories to be repeatedly re-ingested, potentially amplifying duplicates, outdated information, and incorrect source
attribution.

Changes

• Preserve the original user prompt before injecting recalled context.

• Send only the original prompt to CAPTURE_PROMPT.

• Continue injecting recalled memories into the composer for model context.

• Apply the fix consistently to:

  • ChatGPT
  • Claude
  • T3

• Preserve T3’s existing prompt-separation behavior.

• Add shared prompt-preparation logic and regression tests.

Testing

• 4 prompt-capture regression tests pass.
• Browser-extension TypeScript compilation passes.
• Browser-extension production build passes.
• Biome checks pass.
• Full repository suite: 312 passed, 64 skipped, with 21 pre-existing unrelated failures.

Notes

This changes only the content persisted by auto-capture. The prompt sent to the chat platform still includes recalled memories as before.