feat(reports): unified Risk Summary block — one checklist joining the five scattered company-risk signals

[Schlaflied]

Problem

A candidate evaluating one company now gets five independent risk signals, each living in a different place in the report (or outside it entirely):

Background — "Block G" is the Posting Legitimacy assessment in modes/oferta.md: the ghost-job / scam screen that produces the **Legitimacy:** {High Confidence | Proceed with Caution | Suspicious} header line. It predates the issue-first process — introduced directly in 3a636ac (2026-04) and documented in the README via #862. It has since become the anchor that newer risk signals attach to (#1631's flag line lives inside it; #1685 extends it).

Signal	Where it lives today	Origin
Posting Legitimacy tier	Block G + **Legitimacy:** header line	3a636ac
Employment classification warning (contractor-vs-employee language)	A flag line inside Block G	#1631
Culture Screen (pass / caution / fail)	A field inside Block A	#1641
Interview red-flag score	A separate file: interview-prep/{company-slug}-redflags.md	#1233
AI/transformation buzzword–infrastructure mismatch	Block G (in flight)	#1685

Answering the question the candidate actually asks — "is this company safe to join?" — currently requires reading Block A, Block G, and a sidecar file, and mentally joining them. Nothing in the report ever puts these on one screen. A posting can be High Confidence legitimate in the header while the Culture Screen says fail two hundred lines up and a red-flag file next door says HIGH — and no single place shows that contradiction.

Proposal

Add a ## Risk Summary block at the end of every evaluation report (after Block G, before the Machine Summary): one line per signal, fixed order, three states per line — ✅ clear / ⚠️ {finding} / — not evaluated.

## Risk Summary

| Signal | Status |
|--------|--------|
| Posting legitimacy | ✅ High Confidence |
| Employment classification | ⚠️ contractor-style language: "{quoted phrase}" |
| Culture screen | ⚠️ caution — {evidence} |
| Interview red flags | — no interview sessions yet |
| AI claims vs. infrastructure | ✅ consistent |

Rules:

• Aggregation only, zero new judgment. Each line quotes/links the verdict already produced by its source block or file. The summary never re-scores, re-weights, or overrides — if a line is wrong, the fix belongs in the source signal.
• The interview red-flag line is a cross-reference, not a copy: if interview-prep/{company-slug}-redflags.md exists, surface its current score + link; otherwise — no interview sessions yet. This also gives the red-flag detector (feat(interview): company red-flag detector from transcript signal #1233) a consumer inside the report for the first time.
• — not evaluated is a first-class state. Batch-mode reports that skip a signal say so explicitly instead of silently omitting the row, so an all-✅ summary can be trusted.
• Mirror the block into ## Machine Summary as a risk_summary: map so downstream scripts (analyze-patterns.mjs, dashboard) can consume it without re-parsing prose.

Why now (and why this isn't the aggregation anti-pattern)

When this was last discussed, the concern was building an aggregation layer ahead of its consumers. That precondition has flipped: three of the five signals are merged (#1631, #1233, #1641), one is core Block G, and #1685 is in review. All five producers are real; the only missing piece is the join. If #1685 lands after this, its row starts as — not evaluated and activates with the merge — no ordering dependency.

Scope

• modes/oferta.md: add the ## Risk Summary section to the report template + rendering rules (prompt-level).
• modes/interview-redflag.md: one line noting that its output file is cross-referenced by the report's Risk Summary.
• batch/batch-prompt.md: same block, with batch-mode signals marked — not evaluated where applicable.
• Docs: report format description in AGENTS.md/CLAUDE.md (Blocks A–G + Risk Summary).
• No new scripts. Optional follow-up: verify-pipeline.mjs check that reports ≥ the adopting version contain the block.

Happy to submit a PR if this direction looks right.

[santifer]

A unified Risk Summary joining the five scattered signals is a real reader win, @Schlaflied, and the Block G history in the body is appreciated context. Heads up on timing: #1738 lands in territory adjacent to the analytics block merging right now in a fixed order, and Block G surfaces are moving with it. I'll review #1738 once those settle so you don't have to rebase twice.