feat(sdk): improve read_url error diagnostics and html table extraction

[nilaymallikk]

Summary

Closes #1121.

Enhances the read_url tool in @codebuff/sdk and common tool schemas:

• Structured Error Diagnostics: Adds typed error classifications (errorCode such as HTTP_ERROR, BLOCKED_ADDRESS, INVALID_URL, TOO_MANY_REDIRECTS, UNSUPPORTED_CONTENT_TYPE, RESPONSE_TOO_LARGE, TIMEOUT, ABORTED) and forwards HTTP status codes (e.g. 404, 429, 500) so caller agents can differentiate retryable rate-limits from permanent errors.
• HTML Table Extraction: Added formatHtmlTables pre-processor during HTML readable text extraction to transform <table> elements into structured GitHub-Flavored Markdown tables instead of flattening cells into unformatted lines.

Scope

All changes are strictly scoped within:

• sdk/src/tools/read-url.ts
• sdk/src/__tests__/read-url.test.ts
• common/src/tools/params/tool/read-url.ts

Verification

• Passed unit tests: bun test sdk/src/__tests__/read-url.test.ts (21 pass)
• Passed full SDK test suite: bun test sdk/src/__tests__/ (470 pass)
• Passed CI build and packaging: bun run ci
• Passed binary smoke test: bun cli/scripts/smoke-binary.ts cli/bin/freebuff

[codebuff-team]

Good instinct and correctly scoped to sdk/common (not touching any forbidden paths). The errorCode/status additions to read_url are backward-compatible (optional fields) and well-tested for the HTTP_ERROR/UNSUPPORTED_CONTENT_TYPE/INVALID_URL/BLOCKED_ADDRESS/TOO_MANY_REDIRECTS paths.

Two things need fixing before this is portable:

1. Table cell escaping is missing. In formatHtmlTables (read-url.ts ~L200-235), cell text is joined with | but never escapes literal | characters that may appear in the extracted text (e.g. a table cell containing a|b). That will silently corrupt the resulting markdown table structure. Needs cell.replace(/\|/g, '\|') or similar before joining.

2. Error classification via string matching is fragile. isInvalid (L449-452) and isTooLarge (L534-537) key off substrings of thrown error messages ('Only http:// and https:// URLs are supported', 'Response is too large', 'Response exceeded') rather than typed errors from assertUrlAllowed/fetch. If those message strings ever change, the errorCode silently degrades to the wrong classification with no test to catch it. Prefer having assertUrlAllowed throw a typed error (e.g. {code, message}) rather than matching text.

3. Test coverage gap: you added TIMEOUT, ABORTED, and RESPONSE_TOO_LARGE to the error-code union and wrote classification logic for them, but there's no test exercising any of those three paths — exactly the paths most likely to regress given point 2.

Also worth a nested-table caveat: the regex-based <table>...</table> matcher will misbehave on nested tables (rare in practice, but worth a comment/guard).

Address the pipe-escaping bug and add tests for the untested error codes and this becomes straightforward to port.