fix: complete metadata stripping for all channel sentinel patterns (#211)

[AliceLJY]

Summary

Feishu channel metadata was leaking into stored memories because metadata stripping had three coverage gaps. The reporter provided Feishu samples confirming the leak.

Root cause

AUTO_CAPTURE_INBOUND_META_SENTINELS defines 6 sentinel labels, but the three stripping functions only covered a subset:

Sentinel label	stripEnvelopeMetadata	shouldCapture	stripLeadingInboundMetadata
Conversation info	Yes	Yes	Yes (leading only)
Sender	Yes	Yes	Yes (leading only)
Thread starter	Missing	Missing	Yes (leading only)
Replied message	Yes	Missing	Yes (leading only)
Forwarded message context	Missing	Missing	Yes (leading only)
Chat history since last reply	Missing	Missing	Yes (leading only)

Fixes

• Gap 1 (stripEnvelopeMetadata in src/smart-extractor.ts): Added Thread starter, Forwarded message context, and Chat history since last reply to the regex alternation.
• Gap 2 (shouldCapture fallback in index.ts): Expanded metadataPattern regex to cover all 6 sentinel labels, with a parenthetical pattern that handles both (untrusted metadata): and (untrusted, for context): forms.
• Gap 3 (stripLeadingInboundMetadata in index.ts): Rewrote to scan the entire text for sentinel + JSON blocks instead of only stripping from the beginning. Metadata appearing after user content is now removed.

Test plan

• All 12 existing strip-envelope-metadata tests pass
• Full test suite passes (0 failures)
• Manual verification with Feishu samples from the reporter

Fixes #211

🤖 Generated with Claude Code

[AliceLJY]

Closing in favor of #229 which takes a better architectural approach (extracted cleanup module + comprehensive tests). Will follow up on #229 with the remaining gaps.