ADR 0108 — Lint adds to risk

Status: accepted · Date: 2026-11-06

This amends point 3 of ADR 0094 and does not change the base weights of ADR 0008. The coverage it reads is ADR 0107. The formula, with the cases, is in risk.

Context

With an artifact, risk rescaled the base to leave 20% for lint: complexity 0.40, churn 0.256, size 0.144 and lint 0.20. A file with no findings went down, because that 20% weighed zero. A file missing from files was filled with zero and counted as measured. Project health, if used, said the same thing about every file.

That mixes three different facts: the snapshot base, the observed debt, and the failure to compare the tree.

Decision

1. Base B does not change. Complexity 50%, churn 32% and size 18%, renormalized when a signal is missing, normalized by its p95. Gamma stays at 1.7. Folders still aggregate by maximum.

2. Lint is not a fourth signal of that base. When debt counts, it is added before gamma and never subtracts:

   D = errors + 0.25·warnings (info does not count)

   nL = clamp(D / 10, 0, 1)

   raw = min(B + 0.2·nL, 1)

   risk = raw^1.7 on 0..1

   score = 100·raw^1.7 on 0..100

   α = 0.2 and Dsat = 10 are provisional: they are not calibrated. RISK_WEIGHTS_WITH_LINT stays exported as card metadata; the index does not use it.

3. Whoever computes risk passes lintView for each file, copied from its scopeHealth: fresh, counts, coverage (state, eligible, measured, reasons), applicability (applicable, na or not-found) and reason. Project health is not passed. The whole artifact is not passed. With no artifact there is no view: a 0 is not invented.

4. When the file brings lintView, it wins over the legacy number and over options.lint. The legacy number is only a bridge for callers that still have no view.

5. Debt counts with fresh === true and valid counts, even when coverage is unknown, partial or unmeasured: that is an observed minimum, not a closed measurement. fresh === false or null does not penalize. fresh === null is still “there was no comparison”, with reason freshness-unverified.

6. Not applicable and not found do not collapse. na comes from eligible === 0, from not-eligible or no-eligible, or from applicability: "na" with a reason. not-found is a different state: it is not zero and it is not not-applicable.

7. contributions for complexity, churn, size and lint have points, percent and weight. The points that are not null sum to score. weight is the nominal weight, not the percentage or the renormalized share. points null means the signal does not count.

risk without a lint term is the base alone, the same number as before this term.

Consequences

• A map with two files may not move color: the p95 is the smaller one, both saturate at 1, and adding lint cannot pass 1. Findings and a base that is not already at the ceiling are required.
• Zero findings on a current tree add zero. They do not lower everyone else’s risk.
• The debt lens still paints errors + 0.25·warnings. That is not the index term.
• A card that still reads parts.lint.weight sees 0: that weight does not exist in the base breakdown. The frontend closes the contribution row on this branch.
• α and Dsat can change without touching the scopeHealth contract.

Rejected alternatives

• Keep rescaling to 20%. A lint zero lowered files that nobody could call clean.
• Fill absent files entries with 0. Absence and zero debt stopped being different.
• Use project health for every file. A finding in one file tinted the index of the others.
• Add nothing when coverage is unknown. Errors that were seen would be dropped. They count as an observed minimum, and the file is not called clean.

Roadmap

No new phase.

• Done: the formula in @codarch/metrics and the render, MCP and screenshot wiring.
• Out of this delivery: the contribution row on the card. The frontend patches it on this branch.
• Still open: calibrating α and Dsat. They are provisional.

The record of what is done lives in the guide feed.