launchbound 1.1.0

Added

• The overview shows the field, not just the winner. Under the chosen
  configuration it now lists every measured candidate fastest-first, with what
  each costs relative to the fastest. An autotuner's output is not a
  configuration, it is the claim that the configuration is worth choosing,
  and that claim is unreadable without the alternatives: a field inside a
  percent says the tuning did not matter, one spanning 4x says it did.

  The percentage is anchored on the fastest measured candidate rather than
  on the chosen one, deliberately. Where a refused configuration measured
  faster — the case the overview already warns about in bold — anchoring on
  the chosen would bury the number that matters. On the reduce-flip fixture
  it now reads plainly: the disqualified candidate is fastest and the safe
  choice costs 90.5%.

  The list is the head of view 2's, from one shared ordering, so the two
  cannot disagree. Where it does not fit, it truncates and says how many more
  view 2 holds.

Changed

• The overview fills the pane it was given. It drew four lines into a
  twenty-five-row box; the rest was blank.
• termlens 0.5 → 0.6 for the TUI test harness. No source change was
  needed: the only breaking change in 0.6 is GraphicsSeen becoming Clone
  rather than Copy, and this suite asserts on text. What the upgrade buys
  is a fix that matters here — openpty is now retried when the machine is
  briefly out of PTY devices, which macOS is whenever a suite runs one test
  per core.

CI

• The stress workflow gained a flake hunt. The existing gate is unchanged —
  the TUI suite once per OS on every push and pull request. The hunt runs the
  same suite many times over, split across five machines that each use a
  different --test-threads, on dispatch or weekly. Five machines because a
  race that only loses on a slow runner gets five rolls rather than one; five
  concurrencies because that is the axis a PTY suite's faults live on, and
  running one point five times only samples that point harder.
• A published-package check, at release and weekly: cargo install launchbound-cli into a clean directory from crates.io, then real work with
  no GPU, no network and no checkout — enumerating a config space whose
  constraint rules out exactly one of six pairs. It is the only check here
  that can fail without anybody touching the repository, because cargo install resolves dependencies fresh where CI resolves against the lockfile.
• The gate job no longer caches the analyzer. It cached
  cargo-reconverge and reconverge-driver and skipped the install on a hit;
  a cached binary is not evidence that the gate holds against the published
  analyzer. Nothing in this repository's CI or its action caches anything now.
• Squash merges no longer fail the DCO check. GitHub rewrites a
  web-flow squash commit's author email after the sign-off is written, so the
  exact match the check required was impossible by construction. Such commits
  must still carry a sign-off; every other rule is unchanged.