refactor(agent): unify the agent CLIs into one geist binary (+ chat & memory palace)

[geisten]

What

Unifies the scattered agent/chat demo binaries into a single geist binary with subcommands, over one shared engine:

geist <model> [prompt] [-n N]          generate text
geist agent <model> [request] [-n N]   one-shot tool-use agent
geist chat  <model>                    multi-turn chat + memory palace

Before: 6 binaries. After: 3 (geist + the eval_geist / profile_decode dev tools). geist_agent, geist_shell, and geist_chat are gone — folded into geist.

Commits

1. single-source the agent CLI engine — collapse the duplicated backend/model/session/REPL boilerplate into the reusable geist_agent_main engine (agent_main.h), via a tool-builder callback so a tool's ctx can reference the loaded model.
2. merge geist_agent + geist_shell into geist agent — the two standalone demos become one agent subcommand. Kills the "which binary?" footgun: ./geist no longer silently ignores GEIST_FORCE_CALL / GEIST_AGENT_TRACE.
3. geist chat + memory palace as tools — the last standalone binary becomes the chat subcommand, rebuilt on the agent engine.
4. comprehensive agent + chat E2E tests.

Why

One binary, one engine. Each demo previously re-implemented load → session → REPL → cleanup and drifted (tracing in one, force-call wording in another). Folding them in removes the duplication and the drift, and keeps the embedded single-file build lean (the tool code stays behind !GEIST_EMBEDDED_MODEL).

Engine & memory highlights

• agent.h conversation mode — geist_agent_run gains a conversation flag: it keeps the transcript across calls (opens a fresh user turn, folds the answer back) instead of rebuilding each call, so multi-turn chat reuses the agent loop. Fold-back only advances tlen if it fits, so a full transcript is never corrupted. Rebuilding chat on the engine also retires the old hand-rolled chat framing and its stop-marker leak.
• agent_memory.h — remember(text) (title auto-derived from the first line — single-arg so a forced call can lift it) and recall(slug) geist_tools over the mind.h markdown palace. "Search my notes" reuses doc_search over $GEIST_MIND_DIR — no new search code. Both tools are in geist agent and geist chat; the notes index is injected into the system prompt so recall works one-shot.
• geist chat — conversation + full toolset + memory, with reliable /remember, /recall, /notes slash commands (they bypass the model) and --selftest.

Tests — all green locally (model-gated, Gemma 4 E2B-it)

• New test_agent_tools_e2e.c — drives geist agent across every deterministic tool (list_dir, summarize_file, doc_search, remember, recall, plain answer), asserting both the routed tool (via GEIST_AGENT_TRACE) and the effect (stdout / on-disk note).
• New test_chat_session_e2e.c — one scripted geist chat session: multi-turn memory (turn 2 echoes a name from turn 1 — the conversation-mode proof), /remember + /notes, /help, empty line, /recall, /quit, on-disk note + INDEX.md.
• New test_agent_memory_unit.c — remember writes + indexes, recall loads, bad input errors (no model).
• test_chat_e2e.c / test_chat_int.c / test_agent_main_e2e.c retargeted to the new subcommands.
• Network tools stay covered without the internet by the loopback test_webfetch_int + test_websearch_unit.

All 16 agent/chat/memory tests (unit + int + e2e) pass; clean -Werror -Wall -Wextra -Wpedantic build.

Reviewer notes

• geist_agent.c, geist_shell.c, geist_chat.c deleted; mk/common.mk DEMO_SOURCES updated.
• Docs updated: README (agent + chat sections, GIF re-recorded to show geist agent), docs/agent.md, docs/DEPLOY.md, docs/ARCHITECTURE.md, CHANGELOG.
• Known characteristic (documented in the e2e test): a forced recall lifts the whole request as the slug (a slug isn't a path-like locator), so forced recall takes the slug as the request.