[feat](advisor): weakness-averse second opinion + repair 3 live mismatches in the disclosure

[programmerShinobi]

Two things: the second opinion on the recommendation (step 2 of the formula work), and three defects it uncovered in the compensation disclosure already merged in #70 and live in production.

---

Part 1 — the weakness-averse second opinion

I proposed this as a separate button. Wrong twice: a control the user must first discover is the exact discoverability failure fixed elsewhere in this app, and a second warning panel would compete with the compensation disclosure rather than complete it.

It is one line attached to that disclosure, turning "there is a weakness here" into the question that matters — does the decision depend on it?

Microservices scores 2/5 on Cost efficiency, which sits at 13% in your priorities …
Score weakness more heavily and Serverless (FaaS) wins instead — this pick leans on offsetting.

Additive stays the recommendation; a test asserts running this cannot change what rank() returns.

I overclaimed and my own test caught me

The first draft said a geometric mean means "no amount of strength elsewhere fully recovers" a weakness.

12 equal-weight attributes	geometric
one fit of 1 among eleven 5s	4.37
a flat row of 3s	3.00

The spike still wins. Geometric aggregation reduces compensability; it does not remove it. Real non-compensation needs a veto/threshold rule (ELECTRE-style). Two tests pin the accurate claim so it cannot drift back — and PR #70's description has been corrected on GitHub, since it made the same overclaim.

The function was also named nonCompensatoryWinner while its own doc comment, three lines above, said it is not that. Names outlive comments — someone reads it in six months, assumes the property holds, and cites it. Renamed weaknessAverseWinner.

The documented rates are now machine-checked

28.2% and 12.4% are load-bearing — they are the entire argument for disclosing rather than switching aggregator. The frozen-model guards catch an accidental model change, but a deliberate one updates both copies, passes CI, and leaves these numbers quietly wrong. Both are now measured against a seeded sample and asserted exactly (5634/20000, 2484/20000), so any model move fails and names the comments to rewrite. ~250ms.

---

Part 2 — three mismatches in the shipped disclosure

All live in production since #70. None would have been caught by any existing gate.

1 · The same number, rendered two ways

Seven places display weights through roundWeights(). The disclosure was the only one printing the raw float:

	Priorities panel	the disclosure
Cost efficiency	13%	13.2%
Data consistency	11%	11.3%

One attribute, one session, two numbers. For a product whose pitch is that its arithmetic is auditable, that reads as the tool contradicting itself.

2 · It told the reader they had done something they had not

The copy said "which you weighted 13%". In Guided mode — the default — the reader never sets a weight: they answer 14 questions and the model derives them. The app is careful about this everywhere else (prio.derived reads "Worked out from your answers"), so the sentence contradicted the product's own framing. Now "sits at 13% in your priorities" — true whether derived or edited in Expert mode.

3 · The caveat vanished at export

The Advisor warns on screen, but executiveSummary() — the MADR/report text that actually gets circulated and archived — carried anti-patterns and close-call status and not this.

So the warning disappeared at precisely the moment the decision became permanent. Someone reads the record a year later with no idea the pick leaned on offsetting.

Now recorded in both languages, quoting the same rounded weight the screen shows, and caveating the selected option rather than the top-ranked one — a decision record documents the decision that was made, including a deliberate choice against the recommendation.

---

Checked and found clean

I did not stop at the first finding. Every score display already routes through displayScore (RadarPanel and ScenarioCompare looked like exceptions and are not), and isCloseCall has a single definition.

A test premise of mine was wrong, not the code

An early "stays silent" case changed the levels but left Microservices selected while 100% of the weight sat on time-to-market, where it scores 2/5. The caveat fired correctly; the test was wrong. It now uses Monolith and says why in the comment.

Guard fix, not a guard dodge

check-language-purity flagged the new Indonesian string for the English word "other" — from the {other} placeholder. Placeholders are code, not prose. The guard now strips {…} before analysis rather than the string being reworded to slip past it. Re-verified it still bites.

Gates

build clean · lint 0 · 270/270 unit · 15/15 E2E · 7/7 guards · 0 axe in both languages and both themes at 390px · both bundle budgets within limit with no raise.

🤖 Generated with Claude Code