Live Transcript — make it beautiful, paragraph-style, real-time

• Name: right pane calls computeSpeakerDisplays — full "Ryan Snodgrass" on first appearance, "Ryan" after, "Ryan S." only if two Ryans collide. It's imported by TranscriptViewer today but not LiveTranscript — that's the whole fix (~15 lines). Full name stays as the existing hover identity tooltip.
• Color on the name: already correct in SpeakerTurn.tsx:90 (transcriptColorClass(identityKey)). Left pane draws it in plain text only to show what "no wiring" looks like.
• Coalescing: merge adjacent same-speaker segments into one SpeakerTurn — mirror TranscriptViewer's mergeCuesIntoTurns. One avatar + one name per speaker run; segments flow as one <p>. Optional 2 px left rail in the speaker hex.

• Two-Ryan collision: once a second Ryan is present, computeSpeakerDisplays renders "Ryan S." / "Ryan T." on every non-first turn so they never read as one person. First turn of each still shows the full name.
• Recency: newest turn full opacity; older turns opacity:.74. Freshest segment gets a 260 ms fade + 4 px rise (motion-sageox.md, no bounce, honors prefers-reduced-motion).
• Interim caret: the blinking tail is the isFinal=false styling — dead today (Zoom Scribe Batch emits transcript.final only). It lights up automatically if a streaming ASR ever emits transcript.partial; no component change.
• Unknown: unresolved speaker → stable numbered "Speaker 3" in italic muted, still color-keyed on the "unknown" sentinel. Never leaks a raw id.

• Color removed (DDR-013): speaker-palette.ts, the --ox-identity-* tokens, and the de-collision helper are gone. Identity = avatar + name; the fnv1a hash moves to lib/hash.ts for its non-color use (KB constellation shape).

B · Identity color — one token, flips by mode

Each row shows the speaker's name rendered on both surfaces at once — the AA-verified stop per mode. Same FNV-1a slot; humans anchor sage, agents anchor indigo.

C · Ratified people ramp — more chroma (Ryan) + AA both modes Emory formalizes as DDR

Git check: Emory never purpose-tuned this palette — the stops above are the functional ramps' -700/-800, AA but muddy (8 dark earth tones that converge on cream; sage = forest). A dedicated identity ramp spreads the hue wheel and sets a per-mode lightness so speakers separate at a glance. All 16 stops verified ≥4.5:1 (light on cream / dark on crater).

Ratified (Ryan): people identity uses more chroma than the earthed brand set — a different job (telling humans apart, not brand chrome). Anchors: green = human, indigo = agent. All 16 stops WCAG-verified; the lock test (T6) guards them. Emory formalizes as the DDR + --ox-identity-N tokens.

D · Why light is harder — carry identity on the FILL, not the text ★

First-principles: AA on cream forces dark text (L*≈40); the eye discriminates hue poorly at low lightness, so light-mode names converge while dark-mode pastels (8–11:1, high lightness) pop. Fix = redundant encoding (DESIGN.md §11 r3): the solid avatar + a 2px rail are the primary identity channel, name-color is reinforcement. A filled chip isn't AA-as-text-constrained, so it reads hue far better on cream. Both panes below are forced to light mode.

Second lever if still tight: vary lightness across slots (some darker, some lighter, all ≥4.5:1) so lightness becomes a 2nd discriminating channel on top of hue. But the fill+rail does most of the work.

• 3 dedups, low-risk, independent PRs: initials (3→1), timestamp fmt (2→1), paragraph-break (export one). Each removes real duplication without touching render.
• Boundaries: cast-web keeps its own --color-cast-* ramp (intentional, flips with data-mode); don't force the earth palette onto /cast chrome. Don't build the unified <Transcript> facade until a 2nd new surface needs it — the composers above are enough today.