Make response budget preflight failures actionable

[Widthdom]

Summary

• emit stable E028_RESPONSE_BUDGET_TOO_SMALL JSON for object, bounded-envelope, and NDJSON preflight failures
• report raw requested/effective/minimum bytes, explicit uncertainty, the actual 16 MiB ceiling, and actionable increase-or-reduce retry guidance
• preserve pretty-output retry sizing, capture bounded inner stderr into structured errors, and add bilingual docs plus regression coverage for zero/tiny/exact-minimum/Unicode/duplicate-option/shared-consumer cases
• align the pre-existing Issue Standardize terminal records and exit status for truncated streaming output #4561 NDJSON hard-cap regression with the structured E028 preflight contract after CI exposed its stale stdout/stderr expectation
• incorporate latest origin/main at f12e8f8a8, preserving both the main-side query-freshness accounting and this PR's structured response-budget errors

Root cause

Response-size preflight failures bypassed the normal machine error writer and emitted human-readable stderr (or empty stdout). Clamped budgets and reserialized pretty payloads could also recommend retries that were impossible or still too small.

The CI failure was a stale regression assertion: the old Issue #4561 test still expected empty stdout and human-readable stderr for a one-byte NDJSON budget, while the Issue #4909 contract intentionally returns structured E028 JSON on stdout.

Validation

• CI-failing Issue Standardize terminal records and exit status for truncated streaming output #4561 test: 1/1 passed on both net8.0 and net9.0
• conflict-boundary regressions for Issues Standardize terminal records and exit status for truncated streaming output #4561, Represent clean named queries separately from stale queries #4907, and Keep bounded-response preflight failures machine-readable and actionable #4909: 23/23 passed on both net8.0 and net9.0
• dotnet build CodeIndex.sln -c Debug --no-restore -p:UseSharedCompilation=false
• dotnet build CodeIndex.sln -c Release --no-restore -p:UseSharedCompilation=false
• dotnet format CodeIndex.sln --no-restore --verify-no-changes
• dotnet run --project tools/CodeIndex.Changelog -- check (92 fragments)
• final-review regressions: 4/4 passed on both net8.0 and net9.0
• related budget/bounded/issue-drafts/batch-parallel functional set: net8.0 199/199; net9.0 163 passed, 36 runtime skips, 0 failed
• earlier broad related regression set: 356/356 passed on each runtime
• net8.0 full suite: 11,032 passed, 7 skipped, 0 failed
• net9.0 full suite: E028-related tests passed; the suite reproduced the unrelated snapshot-refresh timeout tracked in Stabilize parallel batch snapshot refresh timeout in the net9 full suite #5015, which passed alone
• Codex adversarial review: two rounds completed; six round-1 and four round-2 findings fixed (two-round workflow cap reached)

Documentation and changelog

• bilingual user/developer/test documentation updated
• changelog fragment: changelog.d/unreleased/4909.fixed.md

Fixes #4909

Follow-up: #5015
Follow-up: #5019