docs/site: remove remaining stale voice-loop and pre-persistent-agent copy

[adrianwedd]

Found during the post-#253 roadmap/doc reconciliation on 2026-08-21.

#253 fixed the primary public story, but two canonical/public surfaces still contain pre-current architecture material.

CLAUDE.md — canonical and therefore highest risk

The top architecture section still says the robot "uses a voice loop (Claude / Codex / Ollama)" and later documents:

• Three backends, same pxh.voice_loop core
• bin/run-voice-loop-claude | bin/claude-voice-bridge
• CODEX_CHAT_CMD subprocess routing

claude-voice-bridge was deleted during the resident-only migration. This is not harmless historical prose: every coding agent is instructed to treat CLAUDE.md as the canonical constitution.

Acceptance:

• current overview describes one resident spark-brain, M5 routes, local perception and deterministic authority;
• no deleted cold-Claude bridge is presented as runnable architecture;
• historical voice-loop details either move to an explicitly dated fossil or are deleted;
• retain genuinely load-bearing operational invariants (arecord, GPIO/audio/health/etc.) rather than replacing the file with vague prose;
• add a structural test rejecting known fossil strings from the current/canonical sections.

Homepage — secondary explanatory sections lag #253

site/index.html's hero/current architecture is good, but the older "How SPARK's Brain Works" / FAQ content still includes framing such as:

• "four things running at the same time" / eyes-and-neck "always moving";
• "Brain — always thinking";
• "inner monologue" in the monitoring FAQ;
• the fridge FAQ still explains the inference as "Claude — the LLM generating the inner thoughts" even though routine reflection is now M5;
• the old roadmap remains hardcoded in static HTML (the current roadmap PR replaces the rendered roadmap via site/js/init.js; static fallback should eventually be reconciled rather than carrying two truths).

Acceptance:

• supporting brain/FAQ copy matches current local perception → M5 → selective resident Claude → deterministic policy architecture;
• do not pretend optional/deferred reflection is continuously "always thinking";
• generic model-generated reflection should not be attributed specifically to Claude when M5 owns routine reflection;
• static roadmap source and rendered roadmap converge on one current truth;
• preserve accurate child-friendly explanations and provenance material;
• current public story tests pin these claims.

Scope

Docs/site only unless verification proves a code claim itself is wrong. Do not redesign the site. Prefer deletion and concise canonical truth over adding caveats.

[adrianwedd]

Grooming 2026-08-26 — confirmed unfixed, and the CLAUDE.md half is worse than "stale prose"

Still present verbatim on master 5b9070de:

• CLAUDE.md:81 — "Three backends, same pxh.voice_loop core:"
• CLAUDE.md:86 — a table row pointing at bin/claude-voice-bridge, a file
  that does not exist
• CLAUDE.md:89 — "Override via CODEX_CHAT_CMD"

The file contradicts itself 300 lines apart: CLAUDE.md:380 says
"bin/claude-voice-bridge is deleted rather than deprecated — leaving
fossils executable is how they turn back into architecture," while :86 still
presents it as a runnable launcher in a table of current backends. An agent
reading top-down hits the fossil first.

That is the strongest version of this issue's own argument: it is not harmless
historical prose, it is a canonical document giving two incompatible answers about
what SPARK's cognition substrate is.

Acceptance criteria in the body are unchanged and right, particularly the last
two: keep the load-bearing operational invariants (arecord, GPIO, audio, health)
rather than replacing the file with vague prose, and add a structural test
rejecting known fossil strings — claude-voice-bridge, CODEX_CHAT_CMD,
"Three backends" — from the current sections. Without the test this recurs.

The site half has not been re-verified in this pass.