feat(admin): the graph read-model, skill enumeration, and record joins (issue #54)

[edgehero]

Second slice of issue #54: the data layer under the trigger/flow graph. Everything the graph will draw becomes readable through read-model.mjs, the admin's one I/O funnel. The view itself is a later slice.

Read-model additions

All never-throw, all degrading per folder to a discriminated unreachable, all bounded by the frozen, literal-pinned GRAPH_LIMITS:

• readFolderSkills: enumerates a cron folder's committed skills from the git object store at HEAD, via the worker's own selectEntries/keepOnlyDeclaredSkills over one hardened ls-tree, plus one bounded cat-file per top-level SKILL.md; frontmatter read through the gate's own newly exported aiTriggerAllows. Display-advisory, never a gate decision: DES-AI-TRIGGER-FLOW-GATE's pre-agent-sha truth is untouched, and an unreadable SKILL.md reads as NOT chainable. Sub-skills (.pi/skills/group/sub/SKILL.md) are carried as isSub with their group; they can never be flows.
• readInjectedSkills: a run.skillsDir working-tree listing, advisory on the doctor precedent (host files have no object store). Surfaces the OQ-022 badge fact: an injected skill carrying ai-trigger: allow is a silent no-op.
• cronRunStats: run counts + last outcome per scheduler id, joined on the RAW repeat:<id>:<millis> jobId with the digits-tail disambiguator (the makeFindPreviousRun doctrine) and regex-escaped ids.
• joinRunsToTriggers: forge records joined via the persisted triggerIndex (PR feat(worker): persist trigger attribution on the run record (issue #54) #165), with the OQ-008 range guard: a stale index counts under unattributed, never lands on whatever entry now occupies the row. Index 0 attributes.
• observedChainEdges: parentJobId joins folded per (parentFlow, childFlow, target), same-target only (OQ-009: cross-folder is unrepresentable, so it is never drawn), self-chains included, chainRefused surfaced per parent flow, capped with a truncated flag.
• collectGraphInputs: the one folder/skillsDir dedupe + caps funnel, so the dashboard seam and the CLI subcommand cannot diverge on it.

Also: readTriggers display records now carry the raw triggers-array index (the identity matched.index counts, cron entries and unusable rows included, so a dropped row leaves a hole rather than renumbering every attribution below it), and resolvePaths mirrors the chain caps with defaults imported from the worker, so the graph can never state a cap the worker does not enforce.

New pure module

admin/src/graph-model.mjs (purity source-regex-tested): parseSkillMeta (frontmatter name/description, the gate scanner's block-isolation discipline, values clipped never refused) and findSiblingMentions (the heuristic half of a potential chain edge). Mentions are boundary-matched against the skill-name charset itself rather than \b, because \b calls a hyphen a boundary and fix would fire inside prefix-fix.

Worker enablers

./materialize exports-map subpath (probed in load.test.mjs's fixture), aiTriggerAllows exported, and CHAIN_DEPTH_MAX_DEFAULT/CHAIN_MAX_PER_JOB_DEFAULT hoisted as exported consts. loadConfig behaviour unchanged, pinned by test.

Specs (same PR)

DES-ADMIN-VIA-PI-EXTENSION AMENDED, and the revision row says out loud what the last three dashboard rows certified as unchanged, because this time it did change: a new read-model surface and new fs/git access. The dashboard's source-regex fs ban UNCHANGED, checked (every new spawn and read lives in read-model.mjs). The .log placement boundary UNCHANGED, checked (the folds read .json records only). DES-AI-TRIGGER-FLOW-GATE UNCHANGED, checked.

Tests

Suite in the CI posture: 2100 pass, 0 skipped; node admin/build.mjs builds. Each new function ships the 6-case degradation set (the readStagedPackages model), GRAPH_LIMITS is literal-pinned frozen, the exec fake throws on anything unrouted (the poller-fake doctrine), and the negative claims are explicit: a stale index never misattributes, cross-target pairs are dropped, deny is not dangling, an unread skill is not chainable.