feat: host config layer + drop hostCapabilities surface

[manojp99]

Adds a persistent host-set config layer between the sealed bundle and per-turn argv,
and removes the hostCapabilities surface that has zero read sites today.

Implements two paired designs:

• docs/designs/2026-06-01-drop-host-capabilities.md (Phase 1)
• docs/designs/2026-06-01-host-config-layer-revisit.md (Phase 2)

---

Phase 1 — Drop hostCapabilities (BREAKING)

Removes the write-only hostCapabilities infrastructure across engine, wrappers, schemas, conformance fixtures, and tests. Both v1 capabilities (supports_steering, supports_structured_errors) resolved to "didn't need engine negotiation" — no behavior was gated on them.

Breaking changes (wrapper API):

• wrappers/typescript/src/index.ts: drop SpawnAgentParams.host field; drop exported HostCapabilities + InitializeHostParams types
• wrappers/python/src/amplifier_agent_client/: drop host/host_capabilities params from spawn_agent, assemble_argv, SessionHandleParams
• --host-capabilities argv flag removed from amplifier-agent run — old callers see click UsageError, exit 2

Cross-repo follow-up: the nanoclaw adapter must drop host: { capabilities: ... } from its spawnAgent call. Tracked separately. ← Reviewer action required

---

Phase 2 — Host config layer

Adds --config <path> resolution (flag > $AMPLIFIER_AGENT_CONFIG env > no config tier → bundle defaults), JSON format with json.load, pass-through schema mirroring downstream module configs, layered merge over bundle declarations.

New surface:

• --config <path> argv flag (wires the previously-stub at single_turn.py:406)
• $AMPLIFIER_AGENT_CONFIG env var
• 4 top-level config keys: mcp, approval, provider, allowProtocolSkew

Removed argv surface (BREAKING — subsumed by config):

• --env-allowlist, --env-extra, --allow-protocol-skew
• AMPLIFIER_AGENT_ALLOW_PROTOCOL_SKEW env var

Internal cleanup:

• provider_detect.py deleted (vestigial — provider selection now from config or new bundle.md default_provider: field)
• XDG resolution consolidated through persistence.py (was duplicated in bundle/cache.py and admin/doctor.py)
• amplifier-agent config show extended to report resolved path, source tier, and parsed values — survives parse failures gracefully

New module: src/amplifier_agent_lib/config/{loader,merger}.py + tests/config/

---

Stats

Commits	65
Files changed	68 (+6,654 / −890)
New tests	Config loader (13), merger (5), integration (2) — all passing
ruff check	Clean
Phase 1 success-metric greps	All zero (modulo intentional removal guardrails)
Phase 2 success-metric adversarials	All hard-fail with expected error codes

Tests

• All Phase 1 + Phase 2 tests pass
• TypeScript wrapper: 62/62 pass
• Python wrapper: 48/48 pass
• Conformance (TS + Py): 5/5 each pass
• Loud-failure verification: amplifier-agent run --host-capabilities '{}' "hi" → click UsageError, exit 2 ✓

Pre-existing failures NOT introduced by this work

17 engine-suite failures pre-date this PR and are out of scope. Verified by git stash + re-run against the baseline. Root causes:

• 12 failures: protocol-0.2.0 drift — test_engine.py, test_engine_with_real_bundle.py, test_phase_2_1_exit_gate.py, test_protocol_gen.py hardcode protocolVersion: "0.1.0" but engine speaks 0.2.0 since commit ea51d05
• 3 failures: --mcp-servers rename gap — test_mode_a_v2_envelope.py, test_single_turn_init_params.py use the renamed flag (renamed to --mcp-config-path in ea51d05)
• 2 failures: pyproject.toml wheel-build duplicate-include — tool.hatch.build.targets.wheel.force-include double-counts protocol/schemas/

Recommend a follow-up tracking issue: "protocol-0.2.0 rename drift + wheel-build cleanup" — independent of this PR's scope.

Design decision history

This PR locks decisions D1–D10 of the host config layer revisit. Key choices made during design conversation:

• JSON over YAML — ecosystem consistency, no Norway problem, no safe_load mandate, no wrapper YAML dep, easier library extensibility
• 2-tier resolution (flag > env), no XDG default — programs-first stance avoids silent shared-$HOME collision
• Pass-through to module configs — amplifier-agent invents no vocabulary; mcp:, approval:, provider: mirror their module's schemas
• Strict-by-default validation — unknown top-level keys hard-fail; forward-compat is the host's responsibility via --protocol-version

Supersedes

• Mode A amendment §2.6 D12 (hostCapabilities field) → superseded by Phase 1
• Mode A amendment §3 (argv surface for env-allowlist, env-extra, allow-protocol-skew) → superseded by Phase 2

Follow-up actions (out of scope)

• File nanoclaw repo issue: "Drop host: { capabilities } from adapter spawnAgent call" — text drafted in docs/plans/2026-06-01-drop-host-capabilities-implementation.md H3
• Tracking issue for 17 pre-existing test failures (protocol-0.2.0 rename drift + wheel-build)
• pnpm-lock.yaml has a pnpm-version-drift marker (.: {}) — unrelated, left uncommitted in this PR