feat: narrative front page + interviewer kit (render-only, rule-derived)

[gkanitz]

Context

Slice E of the v0.2 "Meaning" wave. Shared design context:
docs/superpowers/specs/2026-07-04-impact-roadmap-design.md (PR #47).

Recruiters think in roles and stories, not counts. This slice adds a
narrative paragraph describing the shape of the subject's collaboration
(authoring↔reviewing mix and its trajectory) and an interviewer kit — three
probe questions derived from the strongest signals. Everything is
render-layer, derived from already-attested numbers via published,
deterministic rules.

Goal (vertical slice)

Rule functions over the report struct → narrative sentences + interviewer
probes → rendered front-page section + derivation annex, end to end.
No schema change, no fetching.

In scope

• Pure, unit-testable rule functions in render (or a small sub-package),
  each with a stable rule ID.
• Profile-mix rule: r = reviews_given.total / pull_requests.authored;
  r > 1.5 → review-weighted phrasing; r < 0.67 → authoring-weighted;
  otherwise balanced. authored == 0 → the authoring comparison is skipped
  (never a division blow-up or filler sentence).
• Trajectory rule: split the cadence trend buckets into first/second
  half; compute r per half from bucket counts; skip the sentence when either
  half has < 10 combined events. Relative shift > +0.30 → "shifted toward
  reviewing"; < −0.30 → "shifted toward authoring"; otherwise no sentence.
• Interviewer kit: exactly three probes chosen by documented priority:
  (1) deep reviews present → how they give review feedback; (2) rework stats
  present → how they handle revision cycles; (3) time-to-merge present → how
  they scope PRs; (4) cadence trend present → consistency habits;
  (5) language mix present → depth in the top language. Take the first
  three whose data exists; fewer only when fewer exist. Each phrased as a
  question for the interviewer, grounded in the number it cites.
• Derivation annex: the technical annex lists every rule that fired
  (rule ID + one-line explanation), fulfilling the spec's "how this
  narrative was derived" honesty requirement.
• Sentence templates enumerated in testdata, keyed by rule ID.
• Rule semantics changes require an owner comment on this issue; phrasing
  refinement is free as long as goldens and the wordlist tests pass.

Out of scope

• Any schema change or new data. Bands copy (feat: banded context layer — cited typical ranges under every KPI #48 owns it). Grading or
  seniority labels ("senior", "staff") — the narrative describes shape,
  the reader draws conclusions.

Design constraints

• Privacy safeguards catalog (spec): GitHub Action wrapper with Sigstore attestation and pinned-version verification #6 prohibited-strings, GitLab provider adapter with cross-adapter parity tests #7 no-judgment
  copy.
• No rule may reference data outside the report struct.

Acceptance criteria — TDD

Success (write these tests first; all must pass)

1. Rule-function unit tests: profile mix at r = 1.51 (review-weighted),
   0.66 (authoring-weighted), 1.0 (balanced), authored = 0 (skip);
   trajectory at +0.31 (fires), +0.29 (silent), sparse halves (silent).
2. Golden narratives for fixture permutations: review-weighted,
   authoring-weighted, balanced, with/without trajectory, sparse report
   (minimal sections) — each renders the expected sentences and nothing
   else. No placeholder text ever ("N/A", "unknown", empty braces).
3. Interviewer kit: full report → exactly 3 probes in priority order;
   report missing rework and languages → the next-priority probes fill in;
   minimal report → fewer probes, section omitted entirely when zero.
4. Every numeral rendered in narrative/probes also appears in the report
   JSON (extraction test) — the narrative can never invent a number.
5. Derivation annex lists exactly the rule IDs that fired, matching the
   goldens.
6. Wordlist test over narrative + probes: no grading adjectives
   (excellent, strong, weak, slow, fast, top, elite, poor)
   and no seniority labels (junior, senior, staff, principal).
7. Prohibited-strings golden test passes over the new section.
8. Reports without the narrative's inputs still render (graceful omission,
   no errors) — asserted on the sparse golden.

Failure — QA red flags (any one observed → reject the PR)

• A sentence not traceable to a listed rule ID, or an annex listing rules
  that did not fire.
• Invented or derived-but-unattested numbers in copy.
• Grading/seniority vocabulary, or judgment phrasing smuggled into
  templates.
• Placeholder text in any permutation.
• Rule semantics differing from this issue without owner sign-off in the
  issue thread.

Parallel-work contract

• Territory: render/ — new rules file(s), new template partial (new
  numbered section file before 30-collaboration, e.g.
  render/templates/sections/25-narrative.tmpl), testdata goldens.
• Shared, additive-only: render/render.go (wiring),
  render/templates/layout.tmpl (include line).
• Depends on: feat: banded context layer — cited typical ranges under every KPI #48 merged first is recommended (front page reads better
  with context lines present and goldens then settle once), but there is no
  build dependency. Golden-file conflicts with parallel render slices are
  expected — merges are serialized; rebase and regenerate.

Definition of done

TDD skill followed; all success criteria green; go test ./... passes;
PR references this issue.