fix(webllm): cleanRewriteLine should strip "Here are the rewritten bullets:" preamble line

[Vaishnavi1709]

Surfaced by

The eval harness from #65 (PR #149). Llama 3.2 (3B) consistently emits a leading "Here are the rewritten bullets:" line under the terse and examples-led prompt variants. Example from tests/fixtures/rewrite/reports/eval-rewrite-llama-3-2-3b-instruct-…json:

"perBullet": [
  { "index": 0, "text": "Here are the rewritten bullets:", "startsWithActionVerb": false, ... },
  { "index": 1, "text": "Spearheaded marketing task management, ...", ... },
  ...
]

The preamble survives cleanRewriteLine and ends up as outputBullets[0]. That inflates the output bullet count by 1, fails the action-verb check on cell 0, and dilutes every per-criterion aggregate for Llama.

Root cause

src/lib/webllm/post-process.ts has a PROMPT_ECHO_LINES set that drops scaffolding lines like "rewritten bullets:" and "rules:". Llama's exact phrase "Here are the rewritten bullets:" is not in that set, so the line passes through as a normal bullet.

Fix

Two reasonable shapes:

1. Substring match — change the check from set-membership to substring-contains for a short list of known preamble openers ("here are the rewritten", "here is the rewritten", "rewritten bullets:", "original bullets:").
2. Pattern match — regex like /^here (?:are|is) (?:the )?rewritten/i plus the existing exact-match set.

Either works. (1) is simpler; (2) is more precise. Either way, add a unit test in post-process.test.ts that asserts the line is dropped.

Tracking

• Surfaced by the eval reports committed in PR feat(webllm): rewrite-quality eval harness (#65) #149.
• Related: Phase 3: Rewrite-quality eval harness (model + prompt comparison) #65 (the eval harness itself).