Harden release floor, validate-first tools, and list pagination

[secure-ssid]

Summary

Four related hardening fixes, each with regression tests:

1. Release floor can no longer be disabled (scripts/validate_release.py)

• --min-tools accepted 0/negatives, silently turning the tool-catalog floor into a no-op. A _positive_int argparse type now rejects non-integer and < 1 values with exit code 2 (error: argument --min-tools: expected a positive integer, got '0').
• _build_parser() extracted from main() so CLI parsing is unit-testable.

2. Validate-first ordering (mcp_servers/nac.py, ops.py, config.py)

• test_aaa built the Central client (triggering a live OAuth token refresh) before validating required args; invalid calls now return early with no side effects. This also made the release gate's unit-test step pass hermetically in credential-less environments.
• Same reorder applied to the other tools with the pattern: poe_bounce, port_bounce, cable_test, reboot_device, create_webhook. Clients are now built only after validation — and after user confirmation for destructive ops.

3. Double-applied pagination offsets (config.py, monitoring.py)

• Server-paged list tools re-sliced the returned page with the same offset, so any offset>0 page came back empty with a false truncated: False: passpoint profile/identity lists (_config_list_response), list_config_templates, and list_sites (under CENTRALMCP_BOUND_LISTS). They now slice from 0 and report the true offset, matching the existing list_devices pattern.

4. Bounding-contract violations (config.py, tool_router.py)

• list_config_assignments and list_firmware_upgrades returned raw unbounded responses; both now take limit/offset (plus full_list for assignments) via bound_collection_response.
• Router wrappers list_sites/list_devices advertised a full_list parameter the backend tools don't have — FastMCP silently dropped it, returning a truncated page with no marker. Removed.
• Router list_scopes wrapper exposed no paging, so orgs with >100 scopes couldn't page past the first 100; it now forwards limit/offset/full_list.

No callers depended on the old behavior: nothing in the repo passes --min-tools, and the phantom full_list never worked.

Checklist

• I used fake hosts, fake IDs, and redacted payloads in examples, tests, logs, and screenshots.
• I did not commit real credentials, tokens, tenant IDs, customer data, generated indexes, or local MCP config files.
• I updated matching docs when behavior, setup, environment variables, public tool counts, or GitHub Pages content changed. (No docs reference the changed flags/parameters; public tool counts unchanged at 346 — enforced by test_public_docs_tool_counts_match_catalog.)
• I kept user-facing examples on the low-token router path (find_tool, invoke_read_tool, invoke_tool) where possible. (Not applicable — no user-facing examples changed.)
• I ran targeted tests for the changed behavior (validator, NAC AAA, ops validate-first, webhook, and new test_list_output_bounding.py suites).
• I ran uv run python scripts/validate_release.py --skip-rag — 684 unit tests pass, RAG eval auto-skipped (local indexes absent), tool catalog floor satisfied (346 ≥ 204), exit 0.

Security

For exploitable details or accidental credential exposure, do not use this pull request. Follow SECURITY.md.

Contributor notes

See CONTRIBUTING.md for local setup, validation, and no-secret contribution guidance.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AoFhWvujADVzy5GJq4aQ53