feat(scorecard): expose fallback rejection visibility (#1435)

[ozand]

Summary

Closes #1435.

• Reads existing proposer_reject rows with fallback- cycle IDs.
• Adds fallback_rejects, fallback_rejects_by_reason, and fallback_distinct_target_paths to the loop section.
• Preserves fallback_cycles, fallback_successes, fallback_share, and fallback_success_rate terminal-row semantics unchanged.
• Distinguishes a complete readable window with no fallback rows (0 / {}) from an unreadable ledger (unavailable).
• Does not add any of the new metrics to _TARGETS; no ledger schema or fallback-lane behavior changed.

The operator decision that the fallback lane remains enabled is now reviewable through the existing ledger data. Rejection counts are visibility only and should not be interpreted as lane quality without the demand/starvation context (gaps: []).

Verification

• Scorecard focused suite: 71 passed.
• Added recorded-shaped tests for rejection-only, cycle-only, mixed, neither, and unreadable windows.
• Full pytest baseline: four known Windows failures plus three Three class-A window tests fail on main: the file freezes one clock and leaves two real, so every new branch inherits three failures that look like its own #1432 clock-baseline failures on origin/main; those seven are not caused by this PR and are documented here rather than treated as regressions.

[ozand]

Summary

Closes #1435.

• Reads existing proposer_reject rows with fallback- cycle IDs.
• Adds fallback_rejects, fallback_rejects_by_reason, and fallback_distinct_target_paths to the loop section.
• Preserves fallback_cycles, fallback_successes, fallback_share, and fallback_success_rate terminal-row semantics unchanged.
• Distinguishes a complete readable window with no fallback rows (0 / {}) from an unreadable ledger (unavailable).
• Does not add any of the new metrics to _TARGETS; no ledger schema or fallback-lane behavior changed.

The operator decision that the fallback lane remains enabled is now reviewable through the existing ledger data. Rejection counts are visibility only and should not be interpreted as lane quality without the demand/starvation context (gaps: []).

Verification

• Scorecard focused suite: 71 passed.
• Added recorded-shaped tests for rejection-only, cycle-only, mixed, neither, and unreadable windows.
• Full pytest baseline: four known Windows failures plus three Three class-A window tests fail on main: the file freezes one clock and leaves two real, so every new branch inherits three failures that look like its own #1432 clock-baseline failures on origin/main; those seven are not caused by this PR and are documented here rather than treated as regressions.