[Ticket]: Live-validate local MCP model — per-agent injection + env passthrough

Part of #549

## Type
Chore

## Summary
**Live-validate the local-MCP model** shipped in #548: (1) per-agent tool injection (the lean primary must NOT carry the new tool schemas; each subagent must), and (2) opencode merges a server's declared `environment` OVER the inherited process env (so `npx`/PATH still resolve for `memory`/`jira`/`confluence`).

## Intent
Confirm the security + correctness assumptions of ADR-0071/0072/0073 hold on a real opencode client before relying on them.

## Source of truth (links)
- ADR-0071/0072/0073 "live-validate" follow-ups; `docs/opencode-well-known.md`
- Merged: https://github.com/ADORSYS-GIS/ai-helm/pull/548

## Current Behavior
Validated only by `helm template` (render) — runtime behavior unverified on-cluster/client.

## Expected Behavior
A real `opencode auth login` against `ai.camer.digital` shows: primary lacks `memory_*`/`git_*`/`jira_*`/… ; each subagent has exactly its own; and `@memory`/`@atlassian` actually read/write (env passthrough works, no PATH/ENOENT break).

## Acceptance Criteria
- [ ] `opencode mcp list` / agent tool dumps confirm per-agent injection.
- [ ] `@memory` writes to `~/.local/share/opencode/memory.json` (no ENOENT).
- [ ] `@atlassian` authenticates with user env vars (passthrough confirmed).
- [ ] Findings recorded; any gap filed as a fix ticket.

## Out of Scope
- Adding new servers (other tickets).

## Technical Context
Declared `environment` vs inherited env is the key unknown; if opencode REPLACES rather than merges, `npx`/PATH could break for servers that declare env.

## Risks
- If injection/passthrough assumptions are wrong, primary leakage or broken servers.

## Test Plan
- Fresh login on a clean machine; inspect tool sets per agent; exercise memory + atlassian.

## Human accountable owner
@stephane-segning

## AI Usage Declaration
AI drafted this validation ticket; human owns intent + verification.