fix(workspace): Reports tab renders through the shared display_hint renderers, never a raw dump (#2162)

[trinity-ability]

Fixes #2162

This is a disclosure fix, not a cosmetic one

The Workspace Reports tab dumped an agent-authored payload key-for-key to an external client. That is the same class the sibling surface already refuses: agent_page.py deliberately withholds the free-form asks.context because it "has been a credential-leak surface before (canary G-04)". Reports carry an equally free-form, equally agent-authored blob, and shipped it whole to the agent's customer.

The tab now renders through the shared display_hint renderers, with a client-facing summary fallback instead of a <pre>.

What this does NOT fix — stated precisely

Typed renderers narrow what crosses: each reads only its declared keys, so an unexpected sibling key is not rendered. The fallback bounds, humanises and redacts. Neither eliminates the class:

• a key-value summary still names every top-level key;
• a well-shaped markdown / table / kpi / timeline report still renders its values as authored.

The universal control is a boundary-side scrub at the portal serialization edge, deliberately out of scope here and listed as the first follow-up. Please do not read this PR as closing the leak class.

The redactor is defense-in-depth on the fallback path only. It mirrors the prefix set in canary/invariants/g04_*.py and consumes the whole token. That is a limit, not protection — it does nothing for the four typed paths above.

Operator surfaces are behaviourally unchanged

git diff origin/dev...HEAD -- ReportsPanel.vue ReportsPanelFleet.vue is empty. ReportRenderer's new fallbackComponent prop defaults to null and resolves props.fallbackComponent || ReportJson, so operator call sites — which pass nothing — reduce exactly to pre-branch behaviour.

The human-readable fallback is portal-scoped by design. AC #2 asks for a fallback "deliberately stricter than the operator side, because the audience is an external client." A raw dump is a feature when you are debugging your own agent and a defect when the reader is that agent's customer.

The portal override also covers an agent-chosen display_hint: "json" — a valid value in the MCP tool's enum — not merely a shape mismatch. That closes a second route to a raw dump.

The only operator-visible deltas

Three dark-ink meta-text pairs and one bg-blue-500 → bg-status-info-500 swap that resolves to the identical colour (status-info aliases colors.blue).

The dark fix was necessary, not cosmetic. Measured on rendered DOM, those sites failed WCAG AA in dark:

site	before	after	threshold
table thead / footer	3.67	6.99	4.5
KPI tile label	3.04	5.78	4.5

Light is unchanged at 4.83.

Backend surface

Two optional query params (rows_offset, rows_limit) on the existing portal detail route. No new route, gate, db. accessor, table or column ⇒ CLAUDE.md Rule #9 dual-track does not engage — no SQLite migration, no Alembic revision.

Verified mechanically: nothing touched under db/, migrations/, schema.py, tables.py, models.py, main.py; zero @router. changes; zero Depends( changes in Python (the 6 Depends( hits in the diff are all prose in docstrings and markdown explaining why the operator /rows route is JWT-gated).

REPORT_ROWS_PAGE_MAX is imported from models.py:503 where it already lives — never re-typed — so the two page sizes cannot drift with each side's tests pinning its own version.

Auth and scoping are inherited from the existing route. The uniform-404 property (Invariant #8) survives: report_detail returns None on row["agent_name"] != agent_name before any windowing runs, so a foreign report id is still indistinguishable from a nonexistent one. Pinned by test_missing_and_foreign_stay_indistinguishable_when_windowed.

The route gains a rate limiter it did not previously have

services/rate_limiter.py::enforce, key portal_report_detail:{email}:{agent_name}, 60/60s, placed after _require_roster so an unvalidated path param can never mint unbounded limiter keys.

The windowed read re-reads and re-parses the ≤5 MiB blob per page, so the route that exists to cut transfer raises reads. Amplification is therefore strictly better than pre-branch, which had no limit at all.

Residual, stated: fail-open on a Redis outage, and the key is per-agent, so a client rostered on N agents scales linearly.

Verification

• verify-local PASS (--skip-agent; agent stages correctly skipped — no docker/base-image/** change).
• Live both-themes visual pass on this branch's own dev server, all six report shapes — <pre> count 0 in every case.
• Row window end-to-end: "Showing 100 of 260 rows" → Load-more → "Showing 200 of 260". Without rows_limit the response is byte-for-byte back-compatible.
• At 414px a wide table absorbs 658px of overflow inside its own container while the page never scrolls horizontally.
• The operator/client split was demonstrated on the same report simultaneously — operator shows the raw <pre>, portal shows the summary.
• npm run check:tokens → OK, "dark ink ladder holds in 13 swept files".
• Production vite build succeeds.

Tests

• cd src/frontend && npx vitest run → 25 files / 428 tests passed
• pytest tests/unit/test_2162_portal_report_window.py test_1535_report_prompt_guidance.py test_918_report_endpoint.py -q → 32 passed
• Full backend sweep on this branch (measured in WAVE-4): 9980 passed, 3 failed, 17 skipped, 1 xfailed. The 3 are the pre-existing [XPASS(strict)] reds on dev, reproduced on a pristine origin/dev checkout — not from this branch. The +20 over sibling branches is this branch's own new tests.

The 14 backend tests are overwhelmingly non-happy-path: foreign report, missing/foreign indistinguishability, read-failure→404, limit clamping, negative offset, offset past end, non-mutation of the source payload, rate limiting, route param bounds, and FastAPI route buildability.

AC #4 is not machine-guarded — please read this bit

check:tokens' INK_LADDER_SWEPT covers exactly 13 files and includes neither components/reports/ nor components/portal/. CI therefore structurally cannot see the theme work in this PR; it was verified by hand with the measured contrast above.

Suggest adding both directories to the swept list so the fix is guarded rather than re-earned. They are clean now, so it should pass as-is.

Honest note on the raw-colour ratchet

This branch adds 21 raw-gray class occurrences (ReportSummary.vue +14, PortalAgentPage.vue +4, ReportTable.vue +2, ReportKpiTiles.vue +1; ReportTimeline.vue −1 non-gray), which nudges the count of files exceeding raw-color-baseline.json from 19 → 20.

That increase is the arithmetic of the dark-mode fix itself: repairing text-gray-500 to text-gray-500 dark:text-gray-400 necessarily adds a second raw-gray occurrence. The scanner counts occurrences, while the design system's both-themes rule requires adding the dark variant — so the metric moves the wrong way for a change that strictly improves compliance.

No conversion is available: tailwind.config.js defines semantic tokens for status-*, state-*, brand-*, accent-*, action-* — semantic roles only. Neutral ink has no token (gray is deliberately kept as the raw palette), which is exactly why bg-blue-500 → bg-status-info-500 was converted (info has a token) and the gray pairs were not (they have none).

Rebase expectation

PortalAgentPage.vue, workspace-agent-page.md (## Files table) and core-agent.md §5.11 each have a second writer this wave (#2169). Verified disjoint — my PortalAgentPage.vue hunks are all at old line ≥192, #2169's Overview block is 103–191. Resolution is "take both sides".

Follow-ups — listed, deliberately NOT filed

1. Boundary-side G-04 scrub at the portal serialization edge — the only fix that covers table / markdown / kpi / timeline.
2. Operator stores branch on the literal display_hint, so a prefix-resolved table never pages.
3. The operator /rows 400 dead-ends its own UI.
4. scan-raw-colors.mjs is not CI-wired and its baseline is stale (generated 2026-07-31); measured today, 19 files on pristine origin/dev already exceed it. It also has no self-compare mode — only --baseline (write) and --json — so the "may only shrink" ratchet in CLAUDE.md Rule Feature/process engine #10 is currently unenforced by anything.
5. Renderer-set fallthrough attrs stringify a function into a DOM attribute (pre-existing; fix is inheritAttrs: false across the set).

🤖 Generated with Claude Code

[github-actions]

⚠️ Nightly unit-suite check skipped — merge conflict against dev.

Resolve by running git merge dev locally and pushing the result. The next nightly run will re-test once the conflict is gone.

[vybe]

Validated via /validate-pr. Merges clean onto dev after #2171 and #2172; combined result verified locally (27 files / 550 frontend tests, 20/20 test_2162_portal_report_window.py, prod build, check:tokens all pass). Backend reviewed: ownership check precedes windowing so the uniform-404 (Invariant #8) survives, rows_offset/rows_limit bounded at both route and service layer, shallow copy leaves the source payload intact, rate limiter placed after _require_roster. Operator surfaces confirmed byte-identical (empty diff on both panels; fallbackComponent defaults null and resolves to ReportJson). No new endpoint/table/column, so Rule #9 dual-track correctly does not engage. Security scan clean — flagged token patterns are all fake fixtures (AWS's documented AKIAIOSFODNN7EXAMPLE, .test TLD, redaction-test strings).