Sanitize wizard remarks and make remark parsing robust

Problem

The end-of-run "remark" is free-form model output that we extract from the agent's prose by string-matching a [WIZARD-REMARK] marker (AgentOutputSignals.remark() in src/lib/agent/output-signals.ts) and then ship straight to analytics:

const remark = signals.remark();
if (remark) analytics.capture(WIZARD_REMARK_EVENT_NAME, { remark });

Two problems with this:

1. It's not sanitized (security)

The remark is untrusted, model-generated text captured from output and sent to PostHog analytics with no sanitization. It can carry things we don't want to ingest verbatim: secrets/API keys the model echoed from the environment, file paths or PII, prompt-injection payloads, or control/escape characters. We should sanitize before capturing — strip control chars, redact secret-shaped tokens, cap length, and validate it looks like a single reflection line rather than arbitrary content.

2. Parsing it out of prose is flaky

Marker-scanning the agent's output stream is fragile. We've already hit failure modes:

• The model echoes the ask itself instead of a real remark (fixed narrowly in #806 / the REMARK_INSTRUCTION echo check in remark()).
• Only the first line after the marker is captured; multi-line or marker-mid-line output is dropped or truncated.
• Any change to prose formatting can silently break extraction.

As noted in the thread: parsing wizard logs/output for remarks is inherently flaky and should be reworked.

Proposed direction

Move remarks off free-text prose parsing onto a structured channel — e.g. have the agent emit the remark via a dedicated tool call / structured output field instead of an inline marker, so it arrives as a typed value we don't have to regex out of the stream. Then run whatever we receive through a sanitization pass before analytics.capture.

Acceptance criteria

• Remark content is sanitized (control chars stripped, secret-shaped tokens redacted, length-capped) before being captured to analytics.
• Remark delivery no longer depends on regex-scanning prose — it comes through a structured/typed channel.
• Echo-of-the-ask and empty/garbage remarks are reliably rejected.
• Tests cover sanitization and the new extraction path.

Context: found while debugging the Windows skill-install issue; remarks broke (#806) and Vincent flagged the parsing approach as flaky and wanting a rework.

Start with src/lib/agent/output-signals.ts, including AgentOutputSignals.remark() and the REMARK_INSTRUCTION echo check, then trace the analytics.capture call for WIZARD_REMARK_EVENT_NAME. Review the existing remark flow and determine where the structured delivery and sanitization should be tested; done means typed remark delivery, rejected echoes and empty values, and sanitized analytics payloads.