CL-6204: real tool-payload budget estimation + default Ollama adapter

[TheGreatAxios]

Fixes CL-6535

Summary

CL-6204 (the context-compaction work) merged as Done, but two of its
core pieces were inert on main:

• Budget estimator blind to tool payloads. estimateTurnsChars
  measured excerptBlock's human-readable placeholder
  ([tool_result callId], ~20 chars) instead of a turn's real payload.
  10 turns each carrying a 20,000-char tool_result measured as ~160
  chars against a 32,000-char hard limit — overflow detection and
  compaction never fired, and Ollama silently truncated server-side.
• @corbits/ollama-adapter never loaded by default.
  SIDECAR_ADAPTER_MANIFEST defaulted to [], so a default deployment
  ran the built-in OpenAI adapter against Ollama and never sent
  options.num_ctx, even though the seed already writes per-model
  numCtx quirks onto seeded Ollama offerings.

Changes

• estimateTurnsChars now measures real payload size: tool_call
  argument JSON, tool_result content (including media source sizes),
  thinking/citation/safety/code-execution blocks — kept separate from
  the human-readable excerpt used in the recap summary.
• createBudgetedContextCompactor reserves the summary turn's own
  worst-case size out of the budget before selecting kept turns, so
  folded output can't exceed the budget it folded to (previously the
  summary could land on top of an already-full budget).
• SIDECAR_ADAPTER_MANIFEST's default now registers
  @corbits/ollama-adapter for the "ollama" provider key instead of
  an empty manifest. An operator-set value still fully replaces the
  default rather than merging. This also required hoisting
  @corbits/ollama-adapter to the workspace root — @intx/inference's
  dynamic import() of the specifier, invoked from deep inside its own
  node_modules, cannot otherwise resolve a package only linked into
  individual consumers' node_modules.
• Removed workbench-director.ts's dead duplicate director factory
  (defined/workbenchDirectorFactory/buildWorkbenchDirectorRef): it
  never received a contextBudget and had zero non-test call sites.

Follow-ups filed, not fixed here

• context-budget.ts's hard limit reserves no headroom for the system
  prompt or tool definitions, though its header comment claims to.
  Needs the actual system-prompt size threaded through at that layer.

Test plan

• apps/sidecar package test suite: 186 pass, 0 fail
• packages/ollama-adapter package test suite: 14 pass, 0 fail
• apps/sidecar package-scoped tsc --noEmit: clean
• New test reproduces the reviewer's exact repro (10× 20,000-char
  tool_result turns) and proves the honest-overflow path now
  fires
• New test proves a default boot (no SIDECAR_ADAPTER_MANIFEST)
  resolves ollama and that quirks.numCtx reaches
  options.num_ctx in the built request