docs: re-baseline the MCP context claims against tool search; scope /infinite

[tigers1997]

What & why

Split out of #94 at the reviewer's request. Two of the project's own claims had been overtaken by Claude Code and were overstating what the modules buy. Overclaiming is the real risk here: a user who measures the headline number and finds it wrong stops trusting the rest of the docs.

1. The MCP context claim was factually wrong. README claimed per-task profiles "drop a bloated 4-MCP baseline from ~49% context to under 5%", and docs/04 asserted "every MCP tool is a chunk of JSON schema loaded at session start". Tool search defers MCP schemas by default (alwaysLoad: true is the opt-out), so the premise no longer holds.

Measured rather than re-guessed — four local stdio servers advertising twelve tools each (48 total), identical one-turn session on CC 2.1.245:

Session	Prompt tokens	MCP's share
No MCP servers	26,665	—
4 servers, deferred (default)	27,361	+696 (~14/tool)
4 servers, alwaysLoad: true	40,993	+14,328 (~298/tool)

Deferral removes ~95% of the schema cost; the numbers reproduced exactly across runs. The stale claim was also embedded in four shipped templates — worse than in the docs, because it lands in every user's project: check-context/SKILL.md (its budget guardrails actively taught the model a wrong premise), claude-ctx.sh's rationale comment, servers-cookbook.md (which already explained deferral correctly a few sections earlier, so it contradicted itself), and the mcp.minimal.json comment. All corrected. Profiles are now documented for what they genuinely buy: which servers connect — startup time, auth prompts, cold start, and the blast radius --strict-mcp-config enforces.

The dated experiments-memory example keeps its original result with a superseding addendum rather than a rewrite — an experiment log records what was true when it ran, and that's also a better demonstration of the format.

2. /infinite scoped against dynamic workflows. Workflows now do staged, resumable, budgeted fan-out with structured output between stages. The skill opens with a decision table routing that work to a workflow and keeps the one case it's genuinely good at: N variants of one spec into disjoint slots.

Type of change

• feat — new module / skill / feature (minor bump)
• fix — bug fix (patch bump)
• docs — documentation only
• chore — tooling, CI, release plumbing
• refactor — no behavioral change
• BREAKING

Scope

• One logical change. One commit: re-baseline claims that tool search has overtaken.
• Modules affected: mcp, multi-agent, token-efficiency, experiments-memory
• Personas affected: none — no scaffolded file changes name or location.

Tests

• python3 configure.py --check passes locally — verified on this commit alone.
• No fixture applies; this is prose plus template comments.

CHANGELOG

• Added an entry under ## Unreleased.
• I will SHA-anchor the entry after merge.

License & NOTICE

• My contribution is my own work.
• No AGPL-incompatible code.
• No third-party code added.
• templates/discipline-skills/ untouched by this PR.
• LICENSE / NOTICE untouched.

Signing

• All commits are signed.
• Conventional Commits prefix.

I understand

• An automated AI review will run on this PR.
• No merge is possible while any required check is red.
• My contribution rights are described in CONTRIBUTING.md.

---

Fourth of seven stacked PRs — based on fix/currency-corrections (#94).

[claude]

VERDICT: PASS

Clean docs-only PR. The CHANGELOG ## Unreleased entry is present. No Python logic touched, no schema keys added, no third-party code, templates/discipline-skills/ untouched, and the two topics (MCP token-cost rebaseline + /infinite decision table) are tightly related overclaim corrections that fit a single logical change.

[claude]

VERDICT: PASS

Clean docs correction. All changed files are documentation, template content, or template comments — no code logic, no fixture tests, no persona snapshots. The stale MCP token-cost claim is removed consistently across all affected locations (README, docs/04, docs/06, check-context SKILL.md in both the template and the example, claude-ctx.sh, mcp.minimal.json, servers-cookbook.md), and the addendum pattern used in the experiments-memory file is the correct approach for an immutable experiment log. The /infinite decision table is a sensible scope guard. CHANGELOG ## Unreleased entry is present. No third-party code, no AGPL issue, templates/discipline-skills/ untouched, no schema additions.

[claude]

VERDICT: PASS

Documentation-only corrections: MCP context numbers re-measured against CC 2.1.245 (the old "~49% of 100k window" claim predated tool-search deferral), and /infinite repositioned against dynamic workflows. Both changes are factually grounded and internally consistent across all updated files (README, docs/04, docs/06, two SKILL.md templates, claude-ctx.sh, mcp.minimal.json, servers-cookbook.md, and the experiments-memory addendum). CHANGELOG entry present and detailed.

[claude]

VERDICT: COMMENT-ONLY

Blocking

None.

Advisory

Scope (borderline): The PR bundles two distinct corrections — the MCP token-cost re-baselining and the /infinite vs. dynamic-workflows scoping — in a single commit. CONTRIBUTING.md asks for one logical change per PR. Both fixes share the root cause (stale capability claims overtaken by CC improvements) and a single CHANGELOG bullet covers them coherently, so this is a judgment call rather than a hard violation. Splitting is worth considering for future PRs of this shape, but not blocking here.

infinite/SKILL.md flow (nit): After the new opening section ends with "stop and reach for a workflow instead", the next paragraph begins "Then confirm this is a parallelizable fanout task." The word "Then" implies the reader has already passed the decision-table gate, which is the intent, but a reader skimming may find the transition abrupt. A bridging phrase like "If the work fits that narrow case, then confirm…" would make the two paragraphs read more cleanly. (templates/commands/infinite/SKILL.md:215)

Both are advisory only; the docs corrections are accurate, the CHANGELOG entry is present and detailed, no code paths or persona snapshots changed, no AGPL or MIT subtree concerns.