[P1][Tier 1] Refactor context layer: lean entry point + path-scoped instruction routers (progressive disclosure)

[phenning]

Tier 1 — Foundation readiness · Priority: P1

Gap

.github/copilot-instructions.md is a single ~21 KB / 414-line always-loaded document with domain detail inlined (branching, dependency audits, OTEL naming, logging checklists, etc.). There are no .github/instructions/** path-scoped routers (applyTo globs) and no .github/memory/** index. Area conventions are discoverable via the 26 skills, but standing cross-cutting rules that an edit must obey are not area-triggered, and there is no explicit agent-vs-user doc-audience policy.

Investment

• Split the monolith into a lean always-loaded entry point that links out to on-demand detail.
• Add .github/instructions/** with applyTo globs for high-change/high-risk areas (native AOT addons, MessagePack contracts, lscache, extension vs server TS/C#, tests).
• Add a doc-audience/placement policy separating agent-loaded surfaces from user-facing README/docs; ensure every agent context doc is reachable from an entry point, instruction, memory root, or skill (none orphaned).

Acceptance criteria

• Entry point trimmed to orientation + links (no large inlined domain detail).
• Path-scoped applyTo instructions exist for the top change-prone areas and inline only prevention-critical notes.
• Every agent context document is reachable from an agent-loaded surface.

---

Part of epic #3273

[phenning]

✅ Done — implemented and merged.

PR: Merged PR 759084 — Refactor context layer into lean entry point + path-scoped instructions (vs-green, merged to main as 2df5bddf6e).

What shipped:

• .github/copilot-instructions.md trimmed from ~424 to ~230 lines: a lean always-loaded entry point with a Context Layer & Repository Map (discovery index) and a doc-audience/no-orphans policy.
• Eight applyTo-scoped instruction files under .github/instructions/** (native, dotnet-server, typescript, tests, otel, lscache, diagnostics-logging, dependencies) that load only when a matching file is edited.
• .github/instructions/README.md — authoring guide for agents and humans on which surface to use (entry point vs. instruction vs. skill) and how to add new conventions without re-bloating the entry point.
• A "Context layer authoring" review lens in the code-review skill so future instruction changes are held to this model.

Acceptance criteria:

• Entry point trimmed to orientation + links (no large inlined domain detail).
• Path-scoped applyTo instructions exist for the top change-prone areas and inline only prevention-critical notes.
• Every agent context document is reachable from an agent-loaded surface (repository map; no orphans).

Review note: human review caught that the initial applyTo globs used src/native/** / src/test/**, but this repo's code lives under extension/src/** — those were corrected so the instructions actually auto-load, and the map was reduced to a pure discovery index so scopes stay authoritative in each file's frontmatter.