docs: render org procedures inline for agents-target consumers

[shivros]

Summary

Makes every central TechGodHQ procedure reachable to generic AGENTS.md consumers without duplicating source text.

Problem (COD-444): Creed's agents target renders only config: entries into AGENTS.md — skills: entries render to SkillDir outputs, which the agents target lacks. The org constitution pointed at .creed/skills/{architecture,implementation,review}.md, paths that don't exist in Hydra/Iris consumers.

Fix:

• .creed/manifest.yaml: declare the three procedure files as config: entries (rendered inline into AGENTS.md), replacing the skills: declarations. One authoritative source file per procedure is preserved — the files under .creed/skills/ are untouched.
• .creed/config/org.md: pointer paragraph now references the rendered sections ("follow the Architecture Skill below") instead of nonexistent consumer paths.
• AGENTS.md: regenerated — constitution + all three procedures render inline, org layer preceding consumer context deterministically.
• README.md: updated to describe inline rendering and the shipped layered-source consumption model (was describing the removed pointer design and pre-layered status).

Probe: real Hydra layered render

Creed binary built from main afe5dbb (layered sources; the installed v0.3.0 rejects layered manifests). Probe worktree at Hydra origin/main (c524ccc) with the org layer pinned to this branch's head 9ac26ff0c3a374e4ce1ca3aa999c36c8e517f06f:

$ /tmp/creed validate
Validation passed
$ /tmp/creed sync
agents: 1 written, 0 skipped, 0 failed
$ /tmp/creed sync        # idempotency
agents: 0 written, 1 skipped, 0 failed
$ /tmp/creed diff
$ echo $?
0
$ grep -c "\.creed/skills" AGENTS.md
0                       # zero dangling pointers

Rendered consumer AGENTS.md structure (org procedures precede Hydra's local context deterministically):

1:# TechGodHQ Org Context
49:# Architecture Skill
94:# Implementation Skill
150:# Org Review Skill
205:# Hydra Project Context
238:# Development Instructions
272:# Git / PR Rules

Also verified in agent-context itself: validate / sync / second sync (0 written, 1 skipped) / diff exit 0, and the committed AGENTS.md matches the render.

Notes

• Dormant until consumers re-pin: Hydra/Iris pin the org layer to a commit SHA; they see no change (and no breakage) until they move to 9ac26ff+ — that's COD-446/COD-447, both currently blocked on the Creed v0.4.x release (COD-445).
• Silent-override hazard (documented tradeoff): a consumer declaring a config named org/architecture/implementation/review would last-wins-replace the org entry. No current consumer does. Sanctioned by the issue's acceptance criteria.
• Creed's duplicate_source_path validation forbids declaring the same file as both skill and config, so the skills: entries were removed rather than kept alongside.

Closes COD-444

[shivros]

🤖 Automated Review Panel

Reviewer A (glm-5.1)

Verdict: APPROVE

Verification performed (all independent, real execution):

• Located the actual repo at /home/shiv/agent-context (branch runner/COD-444-skills-reachable-config); re-ran gates: creed validate ✓, sync ✓, second sync idempotent (0 written, 1 skipped) ✓, diff exit 0 ✓.
• Built creed from /home/shiv/creed @ afe5dbb and ran a fresh layered consumer probe pinned to the branch head: validate ✓, sync 1 written ✓, second sync idempotent ✓, diff exit 0 ✓, zero .creed/skills references in the rendered consumer AGENTS.md, org + all three skill sections render before the consumer's own section deterministically.
• Confirmed the three skill bodies render verbatim into the consumer AGENTS.md — no truncation, no duplication; .creed/skills/*.md remain the single authoritative source.
• Checked name-collision hazard against real consumers: Hydra/Iris config names (context, development, workflow) don't collide with the new org entries (org, architecture, implementation, review); layered merge semantics (layer order, last-wins by name) verified in internal/adapters/layered/source.go.

Findings (all non-blocking):

1. README.md is now stale — it still described the old pointer design this PR replaces. (Addressed in-PR: README updated in 9ac26ff.)
2. Five H1 headings now appear in one rendered AGENTS.md — cosmetic; creed concatenates with --- separators by design (pre-existing).
3. Silent-override hazard: a consumer declaring a config named architecture/implementation/review/org would last-wins-replace the org procedure. No current consumer does; inherent to the config-entry approach the acceptance criteria explicitly permit.
4. Fix is dormant until consumers re-pin (pinned-layer design, no breakage either way). Noted in the PR body.
5. PR body must include the exact probe commands and observed result — (Addressed: PR body contains the full probe transcript.)

Goal alignment: all acceptance criteria around reachability, no dangling generated pointers, and single-source-per-procedure are met and empirically verified.

---

Reviewer B (glm-5.1)

Verdict: APPROVE

Goal alignment — verified by live re-execution, not just by reading the diff:

• Rebuilt the probe from scratch: fresh Hydra clone at c524ccc, org layer pinned to the branch head, creed built from afe5dbb. Results: validate passed, sync → 1 written, second sync → 0 written, 1 skipped (idempotent), diff exit 0.
• Consumer AGENTS.md contains the full bodies of all three procedures (byte-identical to .creed/skills/*.md sources apart from the --- separators), zero remaining .creed/skills references, and deterministic ordering: Org → Architecture → Implementation → Org Review → Hydra local context.
• In agent-context itself: validate/sync/second-sync/diff all clean; committed AGENTS.md matches the manifest, so the render isn't stale.
• One authoritative source per procedure preserved: files under .creed/skills/ untouched; only their manifest declaration moved from skills: to config:.

Findings (non-blocking):

1. README was stale and contradicted the PR — (Addressed in-PR: 9ac26ff.)
2. Probe commands/results must be in the PR body — (Addressed.)
3. Documented tradeoff: demoting the files from skills: to config: means a future skill-directory target would not emit them as discrete files — inline-only. Sanctioned by the issue for the current agents-only setup.
4. "below" pointers are order-dependent: hold for the agents target (verified), would dangle under a hypothetical future target rendering org.md alone. Same scope caveat as docs: codify public release authority (COD-451) #3.
5. Minor: multiple H1s — pre-existing creed join behavior, not a regression.

Nothing breaks existing functionality: skill sources unchanged, local agents target validates and renders deterministically, consumer manifests with local skills: [] coexist cleanly with the org layer (validated live).