Replace the context roadmap with a problem inventory

[roiguri]

Deletes CONTEXT_HANDLING_PLAN.md and TEST_HARNESS_PLAN.md; adds docs/plans/context/ in their place.

The roadmap had accumulated three rounds of re-baselining, two withdrawn designs, and a ranking resting on a premise that has since drifted — the heartbeat was 83% of input at the 07-16 measurement and is 63% now, because user-scope spend nearly doubled. Two of ten planned items shipped in five weeks.

Deleted rather than archived, on purpose: a stale plan in docs/plans/ costs reader attention every time the directory is opened, and git history is a perfectly good archive.

What replaces them

• PROBLEMS.md — 33 problems in seven clusters, each stated without its proposed solution, and tagged MEASURED / STALE / ASSERTED / RESOLVED so every claim carries its evidence status. Includes a fresh telemetry reading, and an issue map that is the single place issue numbers appear in the file.
• RESEARCH.md — what outlives the plans: the OpenClaw / hermes-agent comparison, how each measurement was actually taken, and the code-level findings discovered while planning.

Two corrections found while compiling this

• Most issue cross-references in the deleted plans did not resolve. They cite Graceful shutdown never runs: SIGTERM kills the process before uvicorn's handler fires #33, Unify slash-command reply formatting across handlers (app-channel CommonMark collapse) #54 and Image memory & persistence #20 for the umbrella, compaction and event-driven heartbeat; in this repository those numbers belong to unrelated issues, two of them closed. Only Context handling is structurally expensive and structurally lossy (umbrella) #18 was correct.
• turns.jsonl contains an unparseable line — a strict json.loads over the file raises. Any reader must parse tolerantly.

Issues brought into line

Same test applied throughout: a title should state what is wrong, such that someone who rejects every proposed fix would still agree it is true.

#18	rewritten — the old body described the pre-gate picture and had been wrong for weeks
#50	reframed around cross-scope visibility, the real problem behind "channel awareness"
#81, #61, #60	retitled from their solutions to their problems
#59	folded into #36 — same failure, same turn_id, split on shippability grounds
#1, #5	kept as standing concepts, with pointers instead of test or gate lists

Note

PR #70 edits CONTEXT_HANDLING_PLAN.md, which this deletes, and will conflict.

https://claude.ai/code/session_01RdXUG9jwvgS2yu4VPjiNer

[roiguri]

Reopening to re-fire the CI event — no workflow run was dispatched for this PR.

[roiguri]

Merging with an admin bypass — GitHub Actions is in a major outage (qcvjkzcs7j74, started 2026-08-06 15:22:49 UTC), so the required checks cannot dispatch. This repo's last successful run was 15:23:23Z, ~30s after the incident opened.

All three CI guards were run locally against a5341d4, the exact commit merging here:

Guard	Result
path-isolation (required)	OK — no module hardcodes a prod state path
channel-agnostic (required)	OK — import boundaries clean
command-replies (not required)	OK — 17 slash-command replies conform

Bypass authorized by the owner.