Docs: record the 0.15.0 work missing from the changelog - #96

Summary

An audit of the 0.15.0 changelog against every pull request merged since v0.14.3 found seven
shipped changes that were never written down, and a Validation paragraph that described the test
suite without ever stating its result. This fills both gaps.

Documentation only — no code, no behaviour change.

What was happening

The 0.15.0 section is the record of this generation, and anyone writing release notes or answering
"what changed?" works from it. Seven changes that shipped were missing from it entirely, so they
would have gone unannounced at release. Separately, the Validation paragraph opened with "The
engine test suite runs on both .NET 8 and .NET 10" — true, but it never said whether it passed,
which is the one thing a reader wants from that paragraph.

What changed

Four missing capabilities (Added):

• Image delivery to models that can actually accept images, with text-only models told plainly
  rather than silently discarding bytes
• Images counted toward the context-window estimate
• Host-supplied agent tools
• On-demand conversation compaction

Three missing planner changes — the significant ones (Changed and Fixed):

• Per-step verification by a second model is gone (Trust the executor by default: acceptance criteria per step and a final test-and-repair phase #88, Remove strict plan verification and make executor-owned completion the only mode #89). A step used to be gated by a
  separate model call deciding whether it was really done; completion now rests on the executor's
  own evidence against acceptance criteria the step carries, with a closing test-and-repair phase.
  This halves the model calls in a plan run and changes what "done" means — and the changelog
  described neither. Remove strict plan verification and make executor-owned completion the only mode #89's removal of the strictPlanVerification escape hatch was also unrecorded.
• The planner stall when a host stops reading progress (Fix planner hangs and preserve work during verification retries #87), and retries reusing already
  gathered evidence instead of rerunning implementation work.

Validation now reports a result, not an activity. States the passing figure on both target
frameworks, and records what the new command-output-sink coverage actually pins — full command
lifecycle, a non-zero exit reading as a failure rather than a kill, a bare cd producing no
report, output past the model's truncation cap still reaching the sink, and a sink that throws on
every call leaving the command running normally.

Scope and risk

Low. One markdown file under docs/. Nothing is compiled from it and no code path reads it.

The risk worth weighing is editorial rather than technical: whether each description matches what
the change actually does. The two most worth a careful read are the acceptance-criteria /
test-and-repair entry and the strict-verification removal entry, because they describe a real
shift in how a plan decides a step is finished — if the wording is off, it will be off in the
release notes too.

Verification

• dotnet test tests/MandoCode.Tests — 733 passed, 0 failed on both net8.0 and net10.0.
  This is the figure now written into the Validation paragraph, so the claim is measured rather
  than carried over.
• Every merged PR from v0.14.3..main (Document the release process in RELEASING.md #74 through Feature: let a host observe shell commands as they run #95) was listed and matched against the
  [Unreleased] section one by one; these seven are the entries with no corresponding text.

Not covered: nothing here is machine-verifiable. Whether each entry accurately describes its
feature is a judgement call that needs a human read.