PE hyperstack: one framework - personas auto-triggered, compiled from manifests

[KailasMahavarkar]

Unifies MCP + skills + roles + personas into one declared Hyperstack framework, and makes personas AUTO-TRIGGER: the compiled bootstrap now carries the full trigger table generated from persona manifests, so hyper engages a persona the moment a trigger matches - the user never has to name it.

The sync problem this solves

personas/<id>/persona.json held engages_when triggers that never reached the runtime context; the bootstrap's persona bullets were hand-maintained prose in skills/hyperstack/SKILL.md - two sources, guaranteed drift, and no deterministic trigger list in front of the agent.

Before                                    After
──────────────────────────────────        ──────────────────────────────────────────
persona.json (triggers)  ──▸ nowhere      persona.json (id, mode, description,
SKILL.md hand bullets    ──▸ bootstrap        triggers) ──▸ compiler ──▸ bootstrap
   (drift between the two)                SKILL.md ──▸ pointer only (no duplication)

Proof of sync: a trigger appended to bro/persona.json appeared in the recompiled bootstrap with zero SKILL.md edits.

Changes

Piece	Intent
context-compiler.ts	compilePersonaLines() generates the Personas layer from loadPersonas() (injectable param keeps the function pure for tests); gates sorted first; markers extended (marketing, reflect, Triggers:)
Personas layer wording	"auto-engaged by trigger, never waiting to be named"; acting defined (inspection/classification only pre-gate); PASS requires evidence - populated fields = NEEDS-INPUT; net-new vs tweak boundary
skills/hyperstack/SKILL.md	critical block: ONE framework, four layers; 1% Rule covers persona triggers; Persona Registry section becomes a pointer to the manifests
persona.schema.json + 4 manifests	description field (compiled into the bootstrap bullet)
bro sidekick format	quote block: word+reading, kanji breakdown, how natives actually use it, simplified usage - never a dense run-on line; no CJK in box-drawing
reflect feature review	reviews features (shipped or planned) as value verdicts (want / don't care / against + the flipping condition), never invented-UI commentary; screens still get the react; panel mode covers features

Pressure-test evidence

hyper agent, given only the compiled bootstrap and "let's add a referral program - build it" (no persona, skill, or gate named):

• auto-engaged product-manager, quoting the exact trigger line ("net-new feature; build request")
• engaged BEFORE any design/build step, citing "the trigger is the contract"
• ran 4 product_manager_* MCP tools, then HELD the gate at NEEDS-INPUT because its own assessments lacked evidence - even though the MCP resolve tool returned a mechanical PASS
• queued the correct post-PASS pipeline (worktree -> blueprint -> designer -> forge-plan -> test-first -> security-review -> ship-gate -> deliver)
• reported 5 wording gaps; the top 4 are closed by this PR (evidence-PASS, "acting" defined, net-new boundary, gates-before-design ordering)

Verification

• bun test: 87 pass, 0 fail (compiler invariant test feeds an explicit minimal manifest)
• tsc --noEmit: clean
• manifest-edit -> bootstrap sync proven end to end
• em-dash sweep on touched files: clean