feat(aside): first-class Aside agentic-browser support

[intertwine]

Summary

Promotes Aside (agentic browser) from an ad-hoc --source cowork impersonation to a first-class observe source, alongside Claude Code / Codex / Grok / Cowork / Hermes.

Key realization: Aside persists a native per-session transcript at
~/.aside/u/<idx>/agents/<agent>/sessions/<date>_<id>/messages.jsonl, but with its own schema (keyed by role not type, epoch-ms timestamps, assistant content = list of text/thinking/toolCall blocks, no per-message UUID). That is why the Cowork/Claude parser ignored it. This PR gives Aside its own parser + count-based cursor (the Codex/Grok strategy).

What is in here

• transcripts/aside.py — native parser: drops thinking, summarizes toolCall against Aside's real lowercase tool names, skips toolResult/system-message, ms to ISO-8601, discovery via u/*/agents/*/sessions/*/messages.jsonl (plus ASIDE_HOME override).
• observe.py — observe_aside_transcript + observe_all_aside (count cursor, rotation-safe). No om backfill wiring, matching the Grok precedent.
• config.py — Config.aside_home + aside_sessions_root.
• cli.py — --source aside (observe enum + --transcript dispatch + source detection + scan loop), an Aside section in om status, --for help text.
• aside/skills/om/SKILL.md — Aside-side warm-start + write-back skill, shipped in the wheel (verified hatchling ships asset trees; uses the native --source aside flow, no hand-built transcript).
• Docs: docs/integrations.md section, README Agent Support row, plans/aside-first-class-integration.md.
• Tests: TestAsideParser, TestAsideObserver (8 new).

Design decisions needing scrutiny

1. Observe-first peer, no installer. Aside has no native hook surface (daemon + skills), so there is no om install --aside; warm start is skill-driven via om context --for aside. Modeled on the manual Hermes path. Right shape, or do we want a thin installer that drops the skill into the Aside agent dir?
2. Count-based cursor (no UUIDs). Same as Codex/Grok. Reset-to-0 when stored index > current message count (rotation/truncation). Attack this.
3. No backfill (Grok parity) — intentional, but call it if you disagree.
4. No version bump. 0.9.0 will batch the larger interim post-0.8.0 work; this rides along.

Adversarial review requested — please try to break:

• Parser edge cases: empty/whitespace lines, invalid JSON, list-typed user content, unknown block types, missing/string/negative/huge timestamp, an assistant turn that is only thinking (should drop) vs only a toolCall (should summarize), gigantic transcripts.
• Privacy: confirm thinking blocks and raw toolResult bodies never reach observations. Do toolCall.arguments summaries risk leaking secrets (e.g. tokens in a bash command)? Should we redact?
• Discovery glob: multi-user (u/1, u/2) and multi-agent layouts; symlinks; non-main agents; does the find_recent 24h window match expectations?
• Cursor/idempotency: interleaved toolResult/system-message records vs the count cursor; re-running observe; concurrent writes while a session is live.
• Cross-platform: is ~/.aside correct on Windows/Linux, or do we need an app-support equivalent like Cowork?
• Source detection: _detect_transcript_source ordering vs other roots.
• Packaging: confirm the skill ships and is not accidentally excluded.

Validation

• ruff check . pass; ruff format --check pass
• pytest tests/test_transcripts.py tests/test_observe.py tests/test_cli_observe.py tests/test_cli_install.py tests/test_cli_doctor.py tests/test_config.py -> 159 passed
• om observe --source aside --dry-run works; om status shows the Aside section (discovers real sessions).

HOLD — do not merge

Opened as draft for peer adversarial review. Do not merge yet, and do not bump the version here (0.9.0 will batch the interim work).