docs(todo): the next-session handoff — ranked build work with file pointers

[linhdmn]

The index said "the three remaining stub tools, the model-driven planner and tier-combo names are the next milestone". Three of those four landed in this session (#33 execution, #34 the reasoner, #35 tier routing), so that note was stale and the real next step was unnamed.

Rewritten as a handoff. Every item was verified against the code, not recalled:

#	Item	Evidence
1	The loop cannot read a file	4 tools, none returns file content. query returns graph elements with a 400-char excerpt plus a rung — it is not a reader. So the reasoner locates parseConfig, learns it is at config.go:41, and cannot look at it — which is exactly why the live demo wrote func parseConfig() {}. write_file's description already promises a "read twin" that does not exist
2	The run has no answer	goal_met stores its synthesis in PartialSynthesis — a field named for the bound case — and the model's rationale sits in a step's why
3	Success criteria are prose	ps.Success appears exactly once in the codebase: as prompt text in reason.go:219. Nothing evaluates it
4	web_search	the last stub; either build the onegw kind=searxng call or delete the tool
5	Guardrail gaps #37 recorded	reply-out unscreened; ~740 ms/call unmetered by BudgetGuard
6	Planning is a rule table	no longer blocking — the chooser carries runs

Each carries the files a doer actually needs to open.

Also in here, because a new session loses time rediscovering it

• The worktree rule, make check as the gate, and the shell quirks that cost time today: a backgrounded server dies without nohup env … &; pkill -f on a name matching your own command line kills your shell; ports 8080 (onegw) and 9699 (leankg MCP) are taken — both verified busy while writing this.
• The one that matters most: every defect this repo shipped lately was a check that could not fail — fix(loop): Categorize matched none of the v1 tool names — every gated run paused on step 1 #25 (a gate that held everything), docs: sync PRD/USAGE/README/todo with the merged code, and make the PRD checker actually run #29 (a checker that could not run), fix(eval): the M6 deploy gate reported "blocked" for days while the tests were green #32 (an eval suite reporting "blocked" while green), feat(guardrail): wire the TypeSafe screen — the containment claim was false #37 (a screen nothing turned on), and TypeSafe guardrail screening experiments + containment proposal #8's checklist ticked on the author's behalf. Verify by breaking the thing you just wrote and watching it fail.

Checks

make check green: PRD OK (12/12, and --selftest 12/12). Docs only — no code changed.

CI shows red — the org's Actions billing limit is still unraised (#31's caveat, recorded in USAGE §2).