Keep E1 upload flags within proved command boundaries

[chrisknvidia]

Fixes #838.

A health-check curl can currently consume -d from a later, independent Docker command and report E1 External Transmission. Limit curl/wget upload-pattern matching to proved command spans so separate LF/CRLF commands and Markdown examples keep their own flags. Preserve multiline uploads and derive locations and fingerprints from original source text. Corrected multi-command spans can change existing fingerprints; baseline generation and rescans are covered.

Reuse the bounded shell tokenizer and retain existing matching for uncertain syntax, including ambiguous same-line semicolon/hash boundaries, Windows shell constructs, malformed input and overlapping nested candidates. Other E1 patterns retain their behavior. The original Docker example can still produce an independent RP1 finding.

Validation on this head: 827 focused tests, 47 installed-wheel CLI checks and five real MCP stdio calls passed. The CLI matrix covers JSON/SARIF, strict exit policy, referenced shell files and baseline generation/suppression. Independent review verified the actual reports and installed package identity. Targeted mypy and source/test Ruff lint and formatting passed. All six hosted checks passed. Additional independent code/security review passed, including 13 native Bash argument-capture comparisons, bounded repeated-fetch timing, and actual installed-console JSON/SARIF scans for redirection, process substitution, parameter expansion and background commands. Native Windows shell execution was not run.

This overlaps #741 in E1 matcher infrastructure. Whichever change lands second must preserve both command boundaries and cumulative matching budgets, including anchored matching. This branch targets current main; combined integration with #741 has not been validated.

[chrisknvidia]

Ready for review at 15b7e324. There are no outstanding review comments or known actionable findings.

All six hosted CI checks have passed.