fix: two-round code review fixes for parser, MCP framing, v0.2 serialization

[superops-team]

Summary

This PR applies fixes identified during a two-round code review of the recent PR #5 (v0.2 upgrade) and PR #6 (MCP server) merges.

Fixes

Parser (boundary bug):

• findFrontmatterEnd had an off-by-one loop bound (i < len(data)-3 → i <= len(data)-3) and did not accept closing --- at EOF without trailing newline, causing valid markdown files without trailing newlines to fail frontmatter parsing.

v0.2 spec conformance:

• ValidateConcept still enforced title as required, contradicting v0.2 spec §4.1 where title is optional. Fixed to derive title from filename when missing (matching parser behavior).

API error handling:

• LoadBundle now explicitly skips reserved filenames (index.md, log.md) before calling ParseConcept, making the skip logic explicit rather than relying on parse errors.

Data loss in SaveKnowledgeBase (critical):

• SaveKnowledgeBase only serialized v0.1 fields when writing concepts to disk, dropping all v0.2 fields (sources, generated, verified, status, stale_after, runtime, parameters, computation, executor, attester, filePath). Fixed to serialize the full v0.2 concept structure.

MCP stdio transport protocol (critical):

• The MCP server used newline-delimited JSON over stdio, but the MCP specification requires LSP-style Content-Length: header framing. Standard MCP clients (Claude Desktop, etc.) could not communicate with the server. Fixed to implement proper Content-Length framing for both reading and writing, with newline-delimited fallback for reading.
• Added stdout flush after writing responses to prevent buffering delays.

Incremental update / delete (v0.2 compatibility):

• After fixing v0.2 serialization, removeKnowledgeFile failed to recognize auto-generated concepts because hasTrustedGeneratedMetadata only checked the legacy generated: true boolean in CustomFields, but v0.2 concepts store generated info in the Generated struct. Fixed to accept both formats.

Archive detection consistency:

• IsArchive was case-sensitive for simple extensions (.zip, .tar) but case-insensitive for compound extensions (.tar.gz), while ExtractArchive internally used case-insensitive checks. Unified to fully case-insensitive detection.
• Fixed .gz suffix check in extractTarGz to also be case-insensitive.

Tests:

• Updated TestValidateConcept_MissingTitle to verify filename-derived title instead of expecting error.
• Updated TestIsArchive/uppercase_ZIP to expect true for case-insensitive detection.
• Updated test_mcp.py to use Content-Length framing matching the fixed MCP server.

Verification

• go build ./... passes
• go vet ./... passes
• go test ./... all packages pass