refactor(mdcode)!: remove entity lookup tools from agent tool derivation

[libei]

Why

entityTools (find_<entity> / lookup_<entity>) was introduced when action parameters were entity references (e.g. Account) and we shipped a standalone ADK agent demo that had no other way to inspect the store.

Since then:

1. Action parameters are pure scalars (feat(mdcode)!: project an action parameter from a field #452): Even projected parameters ({concept: Account, field: accountId}) resolve at load time to a primitive type (Integer) and description; the runtime performs no entity resolution.
2. Agent integration moved to generated Agent Skills (feat(mdcode): generate an Agent Skill from a semantic model #451, demo(mdcode): the commerce demo installs a generated skill, not an agent #453): Generated SKILL.md files emit the physical schema mapping (readableEntities() / boundFields()) so agents can query the store through their native read surface rather than a narrow single-table exact-match SELECT ... LIMIT 51 tool.
3. Keeping entityTools caused SKILL.md to contradict itself: instructionFor() told the agent to "find it with the lookup tools rather than asking for it", while the very next section (## Finding a record) had to walk that back ("This skill offers writes, not reads...").

What changed

• Agent Tool Derivation (src/libts/semantic/runtime/agent_tools.ts):
  • Remove EntityTool, EntityRows, EntityToolOptions, entityTools(), lookupFor(), whyUnreadable(), and runLookup().
  • Simplify ModelTools to {actions, instruction} and CallableTools to {callable, withheld, instruction}.
  • Remove the "find it with the lookup tools" sentence from instructionFor().
  • Keep boundFields() and readableEntities() for skills.ts schema generation.
• CLI (src/tool/commands.ts):
  • Remove printLookupTool() and describedPart() from kcmd agent-tools.
• Tests & Goldens (tests/libts/semantic/runtime/agent_tools.test.ts, tests/libts/semantic/skills.test.ts, *.skill.golden.md):
  • Remove entityTools unit tests and update SKILL.md golden files.
• Documentation (docs/semantic-model/actions.md, skills.md, reference.md, profiles.md):
  • Update Section 7 of actions.md and command references to describe the derived write tools and instruction.

[libei]

The deletion itself is clean. No dangling symbols, distinct()'s comment correctly repointed from the lookup/action collision to the action/action one, and "withheld for one of four reasons" matches whyRefusedWithoutRunning (three) plus noStore (one) exactly. The SKILL.md self-contradiction this set out to fix is genuinely fixed.

I rebuilt the payments model from the YAML blocks on the page and ran kcmd agent-tools against it: the listing pasted into actions.md §7 is byte-identical to the live run. Build clean, 1105 tests pass (68 + 997 + 40).

What's left is orphaned prose — the same failure mode as the reading-judge removal.

---

The demo still advertises the deleted tool

demo/semantic-model/skill/catalog/EntryGroups/commerce_demo/commerce.yaml isn't in this diff, so these can't be inline:

• :156 — IssueCredit's ai_context.instructions still says "Look the order up first if you were given a customer name rather than a number." I ran kcmd agent-tools in demo/semantic-model/skill/ and that sentence comes back at line 10 of the output, so it also lands in the generated references/issue-credit.md. The demo instructs the agent to use a tool this PR deletes.
• :53 — the comment explaining why the model's ai_context stays business-only cites "How to use a lookup tool, and what to do when a write is refused, are properties of the tools ... so the derivation supplies those." The derivation no longer supplies the first half.

Worth noting why these survive a grep: lookup tool at :53 is wrapped across two comment lines, so a line-based search can't match it, and :156 is a paraphrase rather than the term.

Minor

• tests/libts/semantic/runtime/agent_tools.test.ts:57-59 — queryStatus, queryMessage and sessionThrows on FakeStore are now unreachable; nothing assigns them since the lookup describes went. (executeQuery itself may still be needed to satisfy the client shape.) Dead scaffolding reads as live coverage to the next person.
• boundFields stays exported but has no production consumer outside agent_tools.ts, where readableEntities calls it at :572. The description says it's kept "for skills.ts schema generation," but skills.ts imports only readableEntities — so it could go module-private on a change whose point is shrinking this surface.

Pre-existing, not this PR

actions.md's listing says store semantic_skill_demo while the profile block at :264 says database bank. Identical at HEAD~1.

---

Nothing here is a correctness bug in the runtime. The demo instruction is the one I'd fix before merging, since it ships a live pointer to a tool that no longer exists.

[libei]

This now tells the agent it totals through its own read surface, but the derived instruction quoted 39 lines below still says "Never compute a total or a balance yourself; the tools do that." With the lookups gone, no derived tool totals anything.

The old wording ("nothing totals anything on its behalf") was merely in tension with that; this version affirmatively directs the agent to do what the shipped instruction forbids. An agent asked what an order is owed reads the lines, then refuses to sum them and waits on a tool that doesn't exist.

One of the two sentences has to move — either drop "or totals" here, or drop the totals clause from instructionFor.

[libei]

This comment says the derivation owes the agent "what a refused write or a warning means" — but the instruction below also still opens with "Never invent an identifier" and no longer says where to get one, since "find it with the lookup tools" went with the tools.

skills.ts patches that hole for SKILL.md via ## Finding a record ("ask the caller, or read the store directly"). The library path — callableTools(modelTools({runtime})), which §7 "Calling it from code" documents as the ADK/LangChain/MCP route — gets only the bare prohibition, so an agent handed a customer name is told never to invent a key and never told where one comes from.

Carrying the skill's one clause into instructionFor would give both consumers the same answer.

[libei]

This assertion no longer tests the thing the describe is named for. Never invent an identifier is in the always-emitted instruction, not in the read-side section, so this line passes even if readSideSection is deleted outright.

The two assertions below it (## Finding a record, the key has to come from somewhere else) do still cover it, so nothing is actually unguarded — but this line is now carrying no weight.

[libei]

Paragraph wasn't rewrapped after "and an agent's lookups" came out — this line is 51 chars and the next runs to 82, inside an 80-col blockquote.

[libei]

@libei Addressed all review comments in 90a5033:

1. instructionFor() & docs/semantic-model/actions.md: Removed the "Never compute a total or a balance yourself; the tools do that." and "or totals" phrasing, restored the key-source instruction ("When you are given a name or a description where an action wants a key, ask the caller or read the store directly."), and updated the kcmd agent-tools listing (store: my-project/my-instance/bank) and golden files.
2. demo/semantic-model/skill/catalog/EntryGroups/commerce_demo/commerce.yaml: Updated the ai_context comment and IssueCredit instructions so neither refers to a lookup tool.
3. tests/libts/semantic/skills.test.ts & tests/libts/semantic/runtime/agent_tools.test.ts: Updated still says where a key has to come from to assert on the ## Actions write-only section ("This skill offers writes, not reads."), removed unused fields from FakeStore, and made BoundField / boundFields module-private while testing readableEntities directly.
4. docs/semantic-model/profiles.md: Rewrapped the blockquote paragraph to 80 columns.