Tracking: agent-era repositioning — prioritized roadmap

Prioritized outcome of the 2026-07 assessment of matlatl's heuristics against how agents actually build context (grep-first discovery, token budgets, the Open Knowledge Format, and the ETH Zurich context-file results). Full rationale lives in the linked issues.

Governing principle: validate before building. The ETH studies (arXiv:2602.11988, arXiv:2601.20404) showed generated context artifacts can hurt agent performance — matlatl's emitted artifacts are in that class, so evidence comes first, and new analytics wait for it.

Priority order

1. #17 — P13: agent-outcome eval harness. Everything below is hypothesis until this exists.
2. #18 — P14: OKF conformance mode. Time-sensitive positioning: the spec is weeks old and has no tooling; matlatl is a near-drop-in for the empty "OKF health gate" slot. Also reframes the structure heuristics from reader-UX (contested) to format integrity (uncontested).
3. #19 — ADR 0016 scent-rationale correction. Docs-only, an afternoon; cheap credibility insurance.
4. #20 — Hops-from-root metric. Cheapest high-value analytic; machinery (streaming APSP) already exists; the agent-correct replacement for raw in-degree discoverability.
5. #24 — Rescoped 2026-07-19: linked code-path liveness already exists in core (verified: broken-link fires on a dead [x](src/gone.go)). Only the backticked-prose-path scanner remains; demoted below #21/#22 and gated on #17 (new analytic, high FP risk).
6. #21 — Token cost as first-class data. Makes every emitted artifact budget-aware; integer-only, trivially deterministic.
7. #22 — Title/heading quality findings. Where scent genuinely applies to agents (catalog titles, not anchors).
8. #23 — Section self-containedness. Novel but highest false-positive risk; value contingent on get-section usage — gate on #17's evidence.
9. #6 — P12 semantic frontier: deprioritized (see comment there) — in the fix-prompt loop the agent already is the semantic tier; revisit only with #17 link-recovery evidence.

Explicitly not planned

• Removing any existing heuristic — the Info-severity/non-gating design means weak-for-agents signals cost nothing; they stay, reframed as maintainer-lint for the LLM-wiki/OKF loop.
• New navigability scalars or centrality variants before #17 arbitrates the current ones.

🤖 Generated with Claude Code

Start by reading the priority list and the linked issues, especially #17, #18, and the completed #19 and #20, to understand the proposed sequencing and existing evidence. This issue is a roadmap rather than a self-contained implementation task; it is done when the prioritized work is reflected in actionable issues with clear scope and dependencies.