callm v0.3.0

This release fixes stalled calls, dropped input, silently ignored options, and
misleading success results across chat, streaming, model discovery, and raw requests.

Timeouts and input

• All API calls default to a 300-second total timeout. --timeout accepts
  seconds or durations (600, 10m); 0 disables it.
• Added --header-timeout and streaming chat --idle-timeout, both inheriting
  --timeout unless explicitly overridden. Zero disables each selected limit.
• Added independent --stdin-timeout (default 300 seconds) and --no-stdin.
  Pipes are read through EOF on all platforms, including delayed producers;
  cancellation interrupts both input and HTTP waits.
• Bound input files/stdin/buffered responses to 64 MiB and SSE events to 8 MiB.

Correctness and provider support

• Parse full SSE events, reject malformed or truncated streams, and stop Anthropic
  streams immediately at message_stop. HTTP and API errors return failure.
• Honor configured clients and endpoint overrides consistently. Follow Anthropic
  model catalog pagination under one deadline and detect repeated cursors.
• Update Claude defaults to claude-sonnet-4-6 directly and
  anthropic/claude-sonnet-4.6 through Straitly/OpenRouter.
• Convert Anthropic images correctly, preserve top-p, validate thinking budgets,
  and preserve explicit token caps. Unsupported JSON-object requests fail explicitly.
• Apply reasoning display controls to both output modes. Preserve original provider
  JSON and large numbers; --json --stats retains separate stderr statistics.
• Request streaming usage for --stats; handle string costs and unknown usage,
  prices, and context sizes without reporting missing data as zero.
• Accept provider flags before subcommands and honor --stream=false. Reject
  malformed numbers, conflicting flags, and unsupported provider combinations.
  Normalize token limits for OpenAI o-series models.

Compatibility changes

• Missing provider keys no longer fall back to another provider's credentials.
  Use the selected provider's key, CALLM_API_KEY, or an explicit key override.
  Cross-origin redirects are rejected before credentials reach the destination.
• Streaming and reasoning display default on only when stdout is a terminal.
  Reasoning goes to stderr; final answers go to stdout.
• Automation relying on incomplete streams, silently dropped settings, or API
  errors returning success must now handle nonzero exit codes.

Documentation and validation

• Synchronize CLI/subcommand help, README, examples, and the callm agent skill with
  current options, aliases, defaults, and configuration precedence. Add a README/help
  consistency test and documented maintenance rules.
• Add mock regression coverage and 91 local CLI scenarios; improve live-test
  assertions and expose make test-race, enabled in Linux amd64 CI.
• Release archives cover Linux amd64/arm64, macOS amd64/arm64, and Windows amd64,
  with SHA256 checksums. Mock validation does not establish live provider generation
  quality, billing, or account-specific model availability.