feat: add global system prompt and seed AGENTS templates; update proposal to include meta/ and prompts/global/; add project AGENTS template\n\nchore: update AGENTS.md and dev logs; encode governance/propagation rule

AGENTS.md

@@ -2,6 +2,8 @@
 
 Scope: Applies to the entire repository unless overridden by a deeper AGENTS.md.
 
+Note: The canonical system rules used by the agent are mirrored in `prompts/global/system.md` (human) and `prompts/global/system.llm.md` (LLM). Keep AGENTS.md and the system prompt aligned. A seed AGENTS template for generated projects lives at `meta/AGENTS.seed.md` (and `.llm.md`).
+
 ## Collaboration Artifacts
 - Root: `collab/`
 - Subdirectories: `collab/questions/`, `collab/proposals/`, `collab/plan/`
@@ -14,6 +16,11 @@ Scope: Applies to the entire repository unless overridden by a deeper AGENTS.md.
 - Begin with a questions file in `collab/questions/` and iterate there until the agent is satisfied or the user instructs to proceed.
 - Do not create a proposal or plan until the questions have been edited by the user and explicitly approved.
 - After approval, create a concise proposal in `collab/proposals/` and a concrete plan in `collab/plan/`, then implement.
+- One-way progression: once a step (questions → proposal → plan) is approved, do not return to a prior step. Perform any subsequent clarifications/edits within the current step’s file for that round (e.g., continue editing `collab/proposals/NN-*.md` after proposal approval) and only create the next step’s file when moving forward.
+- Artifact edits live in-place: Users/humans will edit the non-LLM `.md` files; the agent reads `.llm.md` and writes both `.md` and `.llm.md` siblings for every collaboration artifact.
+
+### One-Time Exceptions
+- Only bypass the Questions→Proposal→Plan cycle when explicitly directed by the user; log such exceptions in `docs/devlog/`.
 
 ## Logs and Documentation
 - Dev logs: `docs/devlog/DEVLOG_LLM.md` and `docs/devlog/DEVLOG_HUMAN.md` updated each meaningful change.
@@ -42,6 +49,7 @@ Scope: Applies to the entire repository unless overridden by a deeper AGENTS.md.
 ## Token/Context Hygiene
 - Chat minimalism: keep messages short; prefer pointers to files over inline dumps.
 - Use `.llm.md` files for concise, machine-friendly summaries; avoid duplicating full human docs.
+- Read `.llm.md` for intake; do not rely on `.md` for reading. Write both `.md` and `.llm.md` when updating artifacts.
 - No large pastes in chat: never paste large code or markdown; write to files and reference paths/sections.
 - Incremental artifacts: start a new `NN-<subject>.md` per round to avoid long single files.
 - Status style: one-line updates before tool calls; rely on devlog for detail.