feat(workbench): CHAOS-5109 render completeness state and requirement outcomes

[chrisgeo]

Ticket: CHAOS-5109

TEST-EVIDENCE

What this change accomplishes, and for whom: CompletenessPanel renders the two answer-completeness signals the contract already carries but the panel did not show: completeness.state (the server-derived not_derived/complete/partial/degraded verdict) and completeness.outcomes[] (one row per requirement, what it was supposed to contain and what became of it — identity, obligation, outcome, impact, served/declared counts, declared cause, and any refinement history). A reader auditing an answer's trustworthiness can currently see only terminal_status (why the turn stopped); they could not see whether the underlying requirements were actually satisfied, which one was not, or what dimension was affected. The two top-level signals can legitimately disagree on the same result, so the panel shows both, unreconciled; every field on a row is either a visible column or a disclosed one, nothing silently absent.

Executed claim (end to end, from scratch):

• Real install: pnpm install --frozen-lockfile in the lane worktree (node_modules a real directory, not symlinked).
• npx prettier --check . → 0 issues.
• npx eslint . → 0 issues.
• npx tsc --noEmit → 0 errors.
• npx vitest run → 60 files / 977 tests passed, including the full CompletenessPanel.test.tsx and DeterministicAnswerView.test.tsx suites (the pre-existing tests unchanged and green).
• npx next build (Turbopack) → compiled, typechecked, and statically generated all 4 routes with 0 errors.
• CI at tip 061e99784237352a8a997e73a4a1ab0286d28418: every workflow run completed, every check-run success (test, Quality, Build, Unit tests, E2E smoke, Dependency Audit, Gitleaks Secret Scan, run_checks.sh self-test, CodeQL) via paginated API conclusions.
• CodeQL on refs/pull/74/head at the same tip: javascript-typescript=0, python=0, actions=0 results_count via code-scanning/analyses.

Enumerated surface touched:

• src/components/CompletenessPanel.tsx — the panel itself: state badge, requirement-outcomes table, per-row refinement disclosure, empty state.
• src/lib/presentation.ts — two new exhaustive tone switches: completenessStateTone (over not_derived|complete|partial|degraded) and planRequirementOutcomeTone (over satisfied|narrowed|unavailable|not_applicable|not_attempted).
• src/lib/contracts.ts — re-exports PlanRequirementOutcomeRow/RequirementRefinement from the generated contract module (already on the wire at the pinned mirror; not previously named by this repo's own stable-name file).
• No contract/schema/example-fixture change: the mirror already carries state/outcomes/refusal_basis at the pinned SHA; none of the 4 example fixtures under src/contracts/examples/ needed editing.

state domain (all 4 members, executed): every member of not_derived|complete|partial|degraded is rendered in its own test with its own title (state: <value>) and its own tone class.

state	tone class	rendered text
not_derived	badge--neutral	"not derived"
complete	badge--ok	"complete"
partial	badge--warn	"partial"
degraded	badge--bad	"degraded"

outcomes[].outcome domain (all 5 members, executed): every member of satisfied|narrowed|unavailable|not_applicable|not_attempted is rendered in its own test with its own title (outcome: <value>) and its own tone class.

outcome	tone class
satisfied	badge--ok
not_applicable	badge--neutral
narrowed	badge--warn
not_attempted	badge--warn
unavailable	badge--bad

outcomes row-shape domain cells (executed):

• outcomes absent → empty-state line, agrees with state: not_derived.
• outcomes: [] → same empty-state line.
• 3 rows, distinct outcomes → 3 table rows, order-checked against the input array index by index (never re-sorted).
• requirement present → identity renders; obligation present → renders on its own line beside the identity (present exactly when requirement is, per the schema); obligation absent → no such line.
• impact → its own column, distinct text from outcome.
• served/declared → their own column, formatted served / declared.
• a row carrying a declared cause (cause_narrowing/cause_coverage/cause_overrun) → the cause renders in the Reason column; cause_observed: false → renders with a "(defaulted)" suffix, distinct from a reported cause.
• refinements present (2 steps, mixed basis/overrun) → a per-row disclosure names the step count and, once opened, every step's own stage/before/after/basis/overrun/coverage; refinements absent → no disclosure rendered.
• requirement absent → row renders as an honest "—", not dropped from the row count.
• 200 rows (the contract's own @maxItems 200 bound) → all 200 render, 0 truncated.
• terminal_status=complete + state=partial (disagreement fixture) → both top-level badges present and unedited.
• every one of the 4 example fixtures under src/contracts/examples/ carrying a completeness block (enumerated via readdirSync, not hand-listed) renders through the panel with no thrown error.

Class sweep: every field PlanRequirementOutcomeRow/RequirementRefinement carries is either a visible column, a disclosed field, or (for refusal_basis) a deliberately unrendered machine field with a named boundary — none is silently absent from the row. refusal_basis (both the top-level InvestigationResult field and its mirrored completeness.refusal_basis) stays a machine field DeterministicAnswerView deliberately never renders (only the accompanying limitation sentence reaches the reader); this PR does not change that boundary, per the ticket's own instruction to leave it exactly where it is rendered today.

SELF-REVIEW (rule 26, executed against this PR's own tip before TELL): re-read the diff against the four boundary requirements — (a) state badge beside terminal_status, own title, exhaustive tone including a distinct not_derived: present, confirmed via rendered title/class assertions covering all 4 members. (b) outcome rows: identity + obligation + outcome + impact + served/declared + reason, wire order, no dedup, explicit empty state agreeing with not_derived, nothing on the row silently dropped: confirmed — outcomes.map with no sort/filter, index-keyed, every row field either a column or a disclosed field, empty branch keyed off outcomes.length === 0 alone. (c) disagreement between state and terminal_status never reconciled: confirmed — the two badges render from independent switches with no cross-reference between them. (d) refusal_basis untouched: confirmed by grep — zero new references to refusal_basis in this diff. Both new tone switches were re-checked for exhaustiveness by removing a case in a scratch copy and confirming tsc reports a compile error on the resulting non-exhaustive switch, then restored.

RISK-NOTES

• Purely additive to CompletenessPanel; no existing DOM elements were removed, renamed, or restyled — completeness-chip-row, completeness-terminal-reason and all prior test ids are unchanged, and the full pre-existing CompletenessPanel.test.tsx/DeterministicAnswerView.test.tsx suites stay green with zero edits to their existing cases.
• No contract, schema, or example-fixture change — this is a pure consumer/rendering change against fields already on the pinned mirror.
• state/outcomes on every one of today's 4 example fixtures happen to be not_derived/absent (the acr example generator has not yet produced a populated-outcomes example), so the populated-table and refinement-disclosure paths are covered by this PR's own synthetic fixtures rather than a pinned example; the empty-state path is covered by real examples.
• New table markup (<table>/<thead>/<tbody>, scope="col" headers, per-row <details>) follows the same accessible pattern already used by CohortRankingPanel/CoveragePanel elsewhere in this codebase.

[linear-code]

CHAOS-5109