[Burntgogi]

Change

Refs #637; resolves its malformed/repeated placement-input item and its bursts traceback item. MTP draft-store correctness and remote narrow-scan cancellation are not part of this PR.

• parse_n_l is the single placement-admission owner: counts in 0..n_expert, no repeated/overlapping explicit layer ranges, and a nonnegative spread total no larger than eligible capacity. All callers pass the trace's expert count; card_sets only materializes the admitted id prefixes.
• Both affected command lines reject repeated single-value options. window-union --k also rejects duplicate window sizes rather than duplicating report rows.
• bursts translates the path-loaded module's SetError at the reader boundary to its own SetError, preserving the named CLI refusal without global module-cache mutation or catching unrelated errors.
• Existing self-tests retain the numerical replay checks and add placement boundaries/refusals plus real CLI regressions. Help and contributor documentation are updated.

Verification

Base 69b66367; head 7cfab04.

• Before: six malformed/repeated CLI cases exited 0; excessive-count and unknown-layer bursts cases exited 1 with Python tracebacks. After: all 8 scenarios exit 1 with the named option and no traceback.
• Full window-union.py --self-test and router-coverage.py --self-test pass, including the existing static/strata numerical checks and seven retained real-CLI refusal cases.
• Separate smoke: valid empty-card and partial-spread placements retain the expected numerical host-slot results; the burst report gives 16 host slots per four-row window. All 11 single-value option forms across these two CLIs reject repetition by name.
• The parser's zero/full-capacity boundary assertions pass. This is not a claim that every all-card report succeeds: the unchanged report's zero-host denominator crash was reproduced on main and separately filed as window-union: a valid placement with no host selections divides by zero #656. This PR does not change report arithmetic.
• Owning static checks: fmt, check-arch, check-comments, and check-levers passed. As in fix(recipes): retain includes and literals inside multiline attributes #658, the initial concurrent worktree runs interfered with the process-global CPU-guard fixture. After serializing those checks, full check-recipes passed at 2498a35a in 137 s. No skipped units, threshold changes or timeout increases were used.
• No GPU execution claim. Local pre-review cannot run without the Claude CLI; remote review/CI remain separate.

Current-head review correction at 7cfab04

The empty-placement MUST_FIX is addressed: both bursts admission checks use spec is not None, so an explicitly empty --n-l is validated and refused instead of silently becoming an omitted placement. The added real-CLI regression failed before the correction and passes after it. Actual CLI execution now exits 1 naming --n-l, without a traceback; omitted, zero-card and partial-card reports remain byte-for-byte unchanged. Both full replay self-tests pass.

The final owning run at committed 7cfab049 passed 5 / 5 (fmt, check-arch, check-comments, check-levers, full check-recipes), wall 159 s. Check-recipes took 138 s; all original bounds and units remain in place. The omitted/empty distinction is documented.

[github-actions]

Verdict: MUST_FIX: bursts accepts an explicitly empty placement and reports a successful host-only replay.

Must fix

1. Send an empty --n-l through placement validation

tools/ref/router-coverage.py:471 bursts <valid-set> --window 4 --n-l "" passes an empty string to bursts. Both truthiness checks skip placement validation, build empty card sets, and return 0 with the same report as omitting the option. Direct parse_n_l("", layers, n_expert) rejects it with a named error. This is a reachable kind A failure in the placement-validation path this fix PR changes.

Fixing both checks now closes the bypass while the placement-validation tests are being added.

Replace the assignment at line 466 with:

    wu = window_union() if spec is not None else None

Replace the condition at line 471 with:

        if spec is not None:

Add this entry to cases in window-union.py's existing CLI refusal test:

            ("router-coverage.py", ["bursts", cli, "--window", "4", "--n-l", ""], "--n-l"),

Checked

Duplicates: no new functions, types, or constants; no duplicate finding.
Complexity: placement-validation and CLI-refusal branches; no engine hot path changes.
Silent failure: followed the empty placement from CLI parsing through card construction to the successful report.
Tests: the CLI entry point with in-memory trace files reproduces the bypass. Both replacements compile and make empty input return 1 naming --n-l; omitted and valid-placement reports remain identical. All three bursts options still reject repetition. The added refusal cases omit empty input.
Gates: no thresholds or pins moved; the upper-count refusal assertion moved to the parser.
Escape catalog: C1–C4 do not match the diff.

Notes

Dryer and crapper did not run because no Rust file changed. Full self-tests were not rerun during validation.

_{🔨 Shingler, bloomery's review bot: approves a clean pull request and requests changes on a must-fix (how it works)}

[Burntgogi]

Addressed the current MUST_FIX in 7cfab04.

Both placement-admission checks now distinguish omission (None) from an explicitly empty string. The empty value reaches the same parse_n_l validation and named-error boundary as other supplied placements. The existing CLI-refusal fixture includes bursts --window 4 --n-l ""; the contributor documentation states the omitted/empty distinction.

FAIL-first: the added regression failed because the command returned 0 with a host-only report. After the fix, the actual command exits 1 naming --n-l and has no traceback. Separate omitted, zero-card, and partial-card report controls remain byte-for-byte unchanged. Both full replay-tool self-tests pass. The final isolated owning checks at committed 7cfab049 passed 5/5 in 159 s (full check-recipes 138 s), with all original units and limits retained.

[github-actions]

Since the last review: must-fix 1 is fixed at tools/ref/router-coverage.py:471. Explicitly empty --n-l reaches placement validation and returns a named error. The new CLI case covers it.

Verdict: CLEAN: The empty-placement bypass is fixed, and the incremental diff introduces no proven defect.

Checked

Duplicates: 0 new functions, types, or constants; no duplicate finding.

Complexity: two guard substitutions; no added branches or engine hot-path changes.

Silent failure: both placement guards traced. Empty input reaches parse_n_l; its named error is translated to the CLI's SetError and returns 1 before printing a report.

Tests: the added CLI case asserts exit 1, a named --n-l error, and no traceback. Restoring the first old guard would produce a traceback; restoring the second would return a successful report. Either mutation fails the new assertions.

Gates: no gates removed, thresholds changed, or pins moved.

Escape catalog: C1–C4 do not match.

Notes

Dryer and crapper did not run because no Rust file changed. Validation used source inspection; full self-tests were not rerun.

_{🔨 Shingler, bloomery's review bot: approves a clean pull request and requests changes on a must-fix (how it works)}

[juliankang4]

@Burntgogi Thanks for the contribution. The review is done, so I'm merging this.